Radar interference signal classification and identification method based on multi-dimensional multi-domain multi-frame feature extraction

Abstract

The invention discloses a radar interference signal classification and identification method based on multi-dimensional, multi-domain and multi-frame feature extraction. The radar interference signal classification and identification method comprises the implementation steps that a randomized interference identification training framework is built; performing multi-dimensional signal feature cross-domain joint extraction to form a high-dimensional feature vector; constructing a distance-speed dimension motion state mapping matrix; the high-dimensional feature vector and the mapping matrix are fused, and a classification recognition model is trained; and outputting the radiation source interference type label. According to the method, a randomized interference identification training framework is used, an identification sample library covering a continuous parameter space is constructed, multi-dimensional and multi-domain feature information of multi-frame data is comprehensively extracted, a high-dimensional feature vector is synthesized, and a classification model is trained in combination with a distance-speed dimension motion state mapping matrix. Accurate identification of the type of the active interference in a complex electromagnetic environment with suppressing interference and pull-off deception interference is realized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of radar communication, and further relates to a multi-dimensional multi-domain multi-frame feature extraction-based interference signal classification and identification method in the technical field of electronic countermeasure. BACKGROUND

[0002] Radar emitter interference signal type identification refers to, in a complex electromagnetic environment, first analyzing and preprocessing the received emitter signal, transforming the interference signal to multiple transform domains such as time domain, frequency domain, time-frequency domain, polarization domain, extracting some characteristic parameters of the interference signal, selecting interference parameters with obvious differences in characteristics to construct a database, and then training a classifier model through a pre-designed classifier method to achieve effective identification of the style of the received signal.

[0003] Northwest University has disclosed a radar active jamming identification method based on feature extraction and combined classifier in its applied patent document (application number 202310831074.9, authorized announcement number CN 116953628 B). The method extracts contour features of time-frequency images, signal complexity, box dimension, and information dimension features to form a feature vector, and realizes classification and identification of radar active jamming signals through a combined classifier, realizes feature extraction and dimension reduction processing of active jamming signals, has small computational complexity, low implementation difficulty, and can classify and identify multiple active jamming signals. However, this method still has the following deficiencies: it is limited to specific signal-to-noise ratios and other interference parameters, and is only effective within a specific parameter range. When the interference conditions in the scene change dramatically, the classification performance will decrease significantly, and the problem of limited generalization performance of traditional radar emitter interference signal type identification classification methods has not been solved.

[0004] Henan University in its application patent document "radar active jamming recognition method, system, storage device and electronic equipment based on multi-domain information fusion network" (application number 202410857768.4 application publication number CN 118859150 A) discloses a radar active jamming recognition method, system, storage device and electronic medium based on multi-domain information fusion network. The method extracts the time domain real part and imaginary part characteristics of the interference signal and performs short-time Fourier transform time-frequency analysis, uses a multi-domain information fusion network for identification and classification, and can accurately identify the active jamming pattern under low jamming noise ratio conditions without prior knowledge of the interference signal, and has a certain robustness. However, this method still has the following shortcomings: in a complex electromagnetic environment, the drag deception jamming cannot be accurately identified, and the recognition accuracy is low SUMMARY

[0005] The purpose of the present application is to overcome the shortcomings of the prior art, and to provide a radar jamming signal classification and recognition method based on multi-dimensional multi-domain multi-frame feature extraction, which aims to solve the problems of significant decline in classification performance, limited generalization performance, and low recognition accuracy in complex electromagnetic environments with suppression jamming and drag deception jamming when the interference conditions in the scene change dramatically.

[0006] The technical idea for achieving the purpose of the present application is that the present application uses a randomized jamming recognition training framework to construct an identification sample library covering a continuous parameter space, so that the model learns the implicit mapping relationship between environmental parameters and jamming types, simulates a complex electromagnetic environment, and maintains the stability of the recognition model when the interference conditions in the scene change dramatically. This solves the problem of limited generalization performance of traditional radar emitter jamming signal type recognition classification methods. The present application uses the characterization of radar signals in each dimension and each transform domain, comprehensively extracts multi-dimensional multi-domain feature information of multiple frames of data, synthesizes a high-dimensional feature vector, and inputs the classification model combined with the distance-velocity dimension motion state mapping matrix. The high-dimensional feature vector reflects the essential characteristics of the jamming signal in each transform domain, and further combines the distance-velocity dimension motion state mapping matrix to introduce the motion frame information of the jamming target, providing complementary discrimination basis for classification, effectively enhancing the classification ability of the recognition method, and achieving effective classification of typical suppression jamming and drag deception jamming, and completing accurate recognition of the type of active jamming signal.

[0007] To achieve the above purpose, the specific implementation steps of the present application include the following:

[0008] Step 1, build a randomized jamming recognition training framework, and construct an identification sample library covering a continuous parameter space;

[0009] Step 2, multi-dimensional signal feature cross-domain joint extraction, extract the distinguishable features of different emitters in multiple feature domains, and form a high-dimensional feature vector;

[0010] Step 3, monitor the radial distance change rate and echo Doppler frequency, and construct a distance-velocity dimension motion state mapping matrix;

[0011] Step 4, fuse the high-dimensional feature vector and the mapping matrix, and train a classification recognition model;

[0012] Step 5, output the radiation source interference type label.

[0013] Further, the step of building the randomization interference recognition training framework is as follows:

[0014] First, build an environmental parameter covering a continuous interval:

[0015] ;

[0016] wherein, indicates the signal-to-noise ratio, indicates the clutter intensity, indicates the signal bandwidth, indicates the signal carrier frequency, indicates the signal pulse width, all environmental parameters are subject to continuous interval distribution, and the value range is , which together constitute a 5-dimensional continuous environmental parameter space;

[0017] Second, build an interference parameter covering a continuous interval:

[0018] ;

[0019] wherein, indicates the interference type, indicates the jamming-to-noise ratio, indicates the interference bandwidth, indicates the interference frequency, all interference parameters are subject to continuous interval distribution, and the value range is , which together constitute a 4-dimensional continuous interference parameter space.

[0020] Further, the step of building the recognition sample library covering the continuous parameter space is as follows:

[0021] First, based on the randomization interference recognition training framework, select a combination of continuous environmental parameters and interference parameters according to a preset step size;

[0022] Second, generate a corresponding radiation source interference signal based on the selected parameter combination;

[0023] Third, store the generated radiation source interference signal and its corresponding environmental parameters, interference parameters and interference type;

[0024] Fourthly, the samples are arranged in order of parameter size to form a sample library for radar jamming signal identification.

[0025] Further, the multi-dimensional signal feature cross-domain joint extraction refers to extracting distinguishable features of different radiation source interferences in time domain, frequency domain, time-frequency domain and other multi-feature domains through multi-dimensional feature space mapping and correlation analysis to form a high-dimensional feature tensor representation.

[0026] Further, the high-dimensional feature vector is as follows:

[0027] ;

[0028] Among them, represents the time-domain waveform sharpness feature, represents the time-domain waveform skewness feature, represents the frequency-domain waveform sharpness feature, represents the frequency-domain waveform skewness feature, represents the time-domain envelope fluctuation degree feature, and the upper subscript T represents the transpose operation.

[0029] Further, the radial distance change rate is as follows:

[0030] ;

[0031] Among them, and respectively represent the target radial distance extracted in the i th frame and the j th frame, is the total number of monitoring frames set.

[0032] Further, the distance-velocity dimension motion state mapping matrix is as follows:

[0033] ;

[0034] Among them, represents the speed change amount in the i th frame and the j th frame, and respectively represent the target radial velocity extracted in the i th frame and the j th frame, is the total number of monitoring frames set.

[0035] Further, the fusion high-dimensional feature vector and the mapping matrix are as follows:

[0036] ;

[0037] Among them, ​​​​​​a high-dimensional feature vector obtained by cross-domain joint extraction of multi-dimensional signal features, a distance-velocity dimension motion state mapping matrix, and a fusion weight coefficient, and satisfying .

[0038] Further, the step of training the classification recognition model is as follows:

[0039] Firstly, a high-dimensional feature vector obtained by cross-domain joint extraction of multi-dimensional signal features, and a distance-velocity dimension motion state mapping matrix constructed based on multi-frame radar signal data are obtained;

[0040] Secondly, a multi-scale span fusion perception classification model is constructed, the high-dimensional feature vector and the motion state mapping matrix are fused at a feature level according to a preset proportion to obtain fusion features;

[0041] Thirdly, the fusion features are taken as model inputs, the fusion features are mapped layer by layer, and a corresponding interference type prediction result is outputted;

[0042] Fourthly, a loss function is calculated according to a difference between the prediction result outputted by the model and a real interference type label, and the classification recognition model is trained by iteratively updating model parameters until the model converges; the difference calculation loss function is as follows:

[0043] ;

[0044] wherein, the total number of feature scales participating in decision fusion in the multi-scale span fusion perception classification model, the loss weight coefficient corresponding to the i-th scale, the interference type discrimination vector outputted at the i-th scale, the real interference type label of the corresponding sample, the difference measurement function for measuring the deviation between the prediction result and the real label. The multi-scale span fusion perception classification model is composed of an input layer, a hidden layer, and an output layer in cascade;

[0045] The input layer is used for receiving fusion features, the fusion features are formed by fusing a high-dimensional feature vector obtained by cross-domain joint extraction of multi-dimensional signal features and a distance-velocity dimension motion state mapping matrix constructed based on multi-frame radar signal data at a feature level according to a preset proportion, the fusion features are inputted into the model in the form of a two-dimensional feature matrix, wherein each row corresponds to a sample and each column corresponds to a feature;

[0046] The input layer is used for receiving fusion features, the fusion features are formed by fusing a high-dimensional feature vector obtained by cross-domain joint extraction of multi-dimensional signal features and a distance-velocity dimension motion state mapping matrix constructed based on multi-frame radar signal data at a feature level according to a preset proportion, the fusion features are inputted into the model in the form of a two-dimensional feature matrix, wherein each row corresponds to a sample and each column corresponds to a feature; ​

[0047] The hidden layer is used for layer-by-layer feature mapping and multi-scale span feature extraction on the input fusion features. and bias parameter vectors with initial values set as and , representing uniform distribution.

[0048] The output layer is used for classification and discrimination of the multi-scale span features output by the third hidden layer, and the parameters thereof include classification weight parameter matrix and classification bias parameter vector, and the output result is used for representing the discrimination result of each radar jamming type.

[0049] In the model training phase, the sample is input into the multi-scale span fusion perception classification model, sequentially passes through the input layer, the hidden layer and the output layer, and obtains the interference type discrimination vector of the output layer , and the difference between the interference type discrimination vector and the real interference type label vector is measured to calculate the loss function , and then the gradient of the hidden layer model parameters is calculated, and the iteration update is performed on each layer parameter according to the preset update step , so that the loss function is gradually reduced with the increase of the iteration number. The training termination condition is set as: when the difference of the loss functions of adjacent two iterations is less than the preset threshold 0.001, or the iteration number reaches the preset upper limit 100000 times, the training is stopped and the converged model parameters are output.

[0050] Compared with the prior art, the present application has the following advantages:

[0051] First, in view of the problem that the traditional interference recognition model relies on the training set of fixed jam-to-signal ratio (JSR) and preset interference parameters, and has weak generalization ability in actual complex electromagnetic environment, and the recognition effect is not ideal, the present application designs a randomized interference recognition training framework, constructs a recognition sample library covering a continuous parameter space, so that the model learns the implicit mapping relationship between environmental parameters and interference types, can simulate complex electromagnetic environment, significantly reduces the dependence of the model on the fixed jam-to-signal ratio condition, avoids the problem of poor interference recognition effect of the traditional interference recognition classification model after the change of environmental conditions, so that the present application can maintain the stability of the recognition model when the interference conditions in the scene change dramatically, and improves the interference recognition ability of the model under different interference parameters and different environmental parameter combination conditions.

[0052] Secondly, the present application uses the characterization of radar signals in each dimension and each transform domain, comprehensively extracts multi-dimensional and multi-domain feature information of multiple frames of data, synthesizes a high-dimensional feature vector, and combines a distance-velocity dimension motion state mapping matrix to provide complementary discrimination basis for classification, thereby overcoming the problem of the existing technology that the feature dimensions of single-time echo data suppression jamming and deception jamming are incompatible, and the recognition accuracy of a traditional classifier is low in a complex electromagnetic environment. The present application realizes accurate recognition of the type of active jamming in a complex electromagnetic environment with suppression jamming and decoy deception jamming. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 is a flowchart of an embodiment of the present application;

[0054] Figure 2 is a result confusion matrix diagram of simulation experiment classification and recognition of the present application. DETAILED DESCRIPTION

[0055] The present application will be described in further detail below in combination with the drawings and embodiments.

[0056] Reference Figure 1 The specific implementation steps of the embodiment of the present application are described in further detail as follows:

[0057] Step 1: Build a randomized jamming recognition training framework.

[0058] Unlike the typical jamming data training framework of a traditional fixed jamming ratio and preset jamming parameters, the present application randomizes the parameter configuration, simulates the process of dynamic game between both parties in an actual electronic countermeasure scene, and constructs a recognition sample library covering a continuous parameter space, a randomized jamming recognition training framework as follows:

[0059] ;

[0060] Among them, denotes an environment parameter covering a continuous interval, denotes a signal-to-noise ratio, denotes a clutter intensity, denotes a signal bandwidth, denotes a signal carrier frequency, denotes a signal pulse width. All environment parameters are subject to a continuous interval distribution, and the value range is , which together constitute a 5-dimensional continuous environment parameter space . denotes a jamming parameter covering a continuous interval, denotes a jamming type, denotes a jamming-to-noise ratio, denotes a jamming bandwidth, The interference frequency is represented. All interference parameters are subject to continuous interval distribution, and the value range is , which together constitute a 4-dimensional continuous interference parameter space .

[0061] Step 2, multi-dimensional signal feature cross-domain joint extraction.

[0062] Through multi-dimensional feature space mapping and correlation analysis, the distinguishable features of different radiation source interference in time domain, frequency domain, time-frequency domain and other multi-feature domains are extracted, forming a high-dimensional feature tensor representation, which provides a basis for interference recognition.

[0063] Step 2.1, extract time domain waveform features, including time domain kurtosis feature and time domain skewness feature . The time domain kurtosis feature is a fourth-order statistic that can be used to represent the sharpness of the time domain waveform distribution of the active interference signal; the time domain skewness feature is a third-order statistic that is usually used to describe the asymmetry of the time domain waveform distribution.

[0064] Step 2.2, extract frequency domain waveform features, including frequency domain kurtosis feature and frequency domain skewness feature . The frequency domain kurtosis feature is a fourth-order statistic that can be used to represent the sharpness of the frequency domain spectrum line distribution of the active interference signal, which describes the kurtosis of the frequency spectrum distribution, i.e. the steepness or flatness of the frequency spectrum distribution; the frequency domain skewness feature is a third-order statistic that can be used to represent the asymmetry of the frequency domain spectrum line distribution of the active interference signal, which reflects the asymmetry of the signal spectrum.

[0065] Step 2.3, extract the time domain envelope fluctuation degree feature, which is a statistical feature quantity used to represent the degree of fluctuation of the signal envelope, calculated by the ratio of the variance of the signal discrete sequence to the square of the mean value.

[0066] Step 2.4, splice multiple signal statistical features to form a high-dimensional feature vector containing multi-dimensional information, which can comprehensively represent the characteristics of the signal in different dimensions and different transform domains: .

[0067] Step 3, construct a distance-velocity dimension motion state mapping matrix.

[0068] The embodiment of the application comprehensively considers the motion characteristics and maneuvering characteristics of the target, monitors the information such as the radial distance change rate and echo Doppler frequency, and constructs serialized data. That is, the distance information and speed information of the target in different frames are presented in a two-dimensional motion state mapping matrix, the matrix is statistically processed, the motion state of the target is analyzed, and a basis is provided for interference identification.

[0069] Step 3.1, processing the received multi-frame signals according to the preset frame interval Selecting a target frame signal, performing pulse compression processing on the signal of the selected frame in the distance dimension to focus the dispersed signal energy, and performing spectrum analysis processing in the speed dimension, and screening out the moving target through Doppler filtering, so as to extract the target radial distance and radial speed (i.e. the speed component corresponding to the Doppler frequency) of the selected frame, and constructing the distance-speed data of the continuous N frame intervals into a two-dimensional frame sequence matrix :

[0070] ;

[0071] At this time, the distance information and speed information of the target are presented in the form of two-dimensional mapping, the two-dimensional sequence matrix constitutes a distance-speed joint mapping space, and each data point represents the distance and speed state of the target at time.

[0072] Step 3.2, calculating the distance change amount and speed change amount of the adjacent interval frames to generate the following dynamic change sequence :

[0073] .

[0074] Wherein, represents the distance change amount of the adjacent interval frames , represents the speed change amount of the adjacent interval frames .

[0075] The dynamic change sequence maps the dynamic change trend of the motion state of the target, and the mean value and standard deviation of ΔR and ΔV in the window are calculated in real time through the sliding window statistical technology (window length M≤N), as the quantitative index of the stability of the motion state.

[0076] Step 3.3, setting a distance dimension mutation threshold based on the prior motion model of the target and the electromagnetic environment characteristics as follows:

[0077] ;

[0078] Wherein, This is an empirical coefficient; in the embodiments of the present invention, this coefficient is taken as 0.8. The maximum allowable distance change is set, and a distance dimension anomaly marker is triggered when the statistical characteristics of the distance change exceed the threshold.

[0079] Step 3.4, set the velocity dimension mutation threshold. for:

[0080] ;

[0081] in, This is an empirical coefficient; in the embodiments of the present invention, this coefficient is taken as 0.8. The maximum permissible velocity change is determined by triggering a velocity dimension anomaly flag when the statistical characteristics of the velocity change exceed a threshold. The threshold parameter is calibrated using Monte Carlo simulation or measured data.

[0082] Step 4: Combine high-dimensional feature vectors and mapping matrices to train a classification and recognition model.

[0083] Based on the high-dimensional feature vector synthesized by cross-domain joint extraction of multi-dimensional signal features and the distance-velocity motion state mapping matrix obtained from the acquisition of multiple frames of signal data, a multi-scale cross-fusion perception classification model is built. The obtained high-dimensional feature vector and the motion state mapping matrix features are fused at the feature level according to a specific ratio. Finally, the prediction results are linearly weighted and summed through decision fusion to obtain the fusion result.

[0084] The input of the multi-scale span fusion perceptual classification model is a two-dimensional matrix, where each row represents a sample and each column corresponds to a feature dimension; the output is the output label, which is generally a one-dimensional vector, and each element in the vector corresponds to the classification result of a sample, i.e., the identified interference type.

[0085] The network structure consists of a cascaded input layer, hidden layers, and output layer. The input layer receives a two-dimensional sample matrix, and the hidden layers output fused features layer by layer.

[0086] ;

[0087] in, Indicates the first The first neuron and the second The connection weights of each neuron Indicates the first The output value of each neuron Indicates the first Bias of each neuron.

[0088] Repeat the above steps until the data is passed to the output layer to obtain the final output predicted label. .

[0089] Finally, the loss function is calculated By minimizing the loss function, the gradient of the weight and bias is updated layer by layer, so that the model prediction result gradually approaches the true label.

[0090] Step 5, identify the radiation source interference type.

[0091] After the model training is completed, the high-dimensional feature vector obtained in step 3 and the motion state mapping matrix obtained in step 4 are input into the classification model to realize stable and fine classification of the radiation source interference signal.

[0092] The effect of the present application can be further proved by the following simulation.

[0093] 1. Simulation experiment conditions.

[0094] The hardware platform of the simulation experiment of the present application is: the processor is Intel i7 9750H CPU, the main frequency is 2.6GHz, and the memory is 16GB.

[0095] The software platform of the simulation experiment of the present application is: Windows 10 operating system and Matlab R2024a.

[0096] The simulation experiment of the present application is a simulation of type identification of a test set of radiation source interference signals simulating a complex electromagnetic environment. The test set setting includes environmental parameters and interference parameters, as shown in Table 1.

[0097] Table 1 Simulation parameter table

[0098]

[0099] The signal-to-noise ratio is set to-10~10dB, the clutter intensity is 0~3, the signal bandwidth is 5~20MHz, the signal carrier frequency is 10~18GHz, the signal pulse width is 20~100us, the interference type is radio frequency noise interference, noise amplitude modulation interference, noise frequency modulation interference, noise phase modulation interference, range decoy interference, speed decoy interference and joint decoy interference, the interference-to-noise ratio is 20~40dB, the interference bandwidth is 5~200MHz, and the interference frequency is 10~18GHz.

[0100] 2. Simulation content and result analysis.

[0101] The simulation experiment of the present application is a simulation of type identification of a test set of radiation source interference signals simulating a complex electromagnetic environment by using the method of the present application, and the identification result confusion matrix diagram is as shown in Figure 2 .

[0102] The test set of the simulation experiment of the present application is composed of 70000 groups of interference signals generated by random parameters in the parameter range of 7 interference types and parameters in table 1 of the simulation parameters simulating a complex electromagnetic environment.

[0103] The effects of the present application are further described below in combination with Figure 2 the simulation diagram.

[0104] Figure 2 The abscissa axis in the simulation diagram represents the predicted labels, including 7 interference types of frequency noise interference, noise amplitude modulation interference, noise frequency modulation interference, noise phase modulation interference, range deception interference, velocity deception interference and joint deception interference, the ordinate axis represents the real labels, which are consistent with the predicted labels, each row and each column corresponds to an interference type, and the values in the matrix represent the proportion of being judged as a certain predicted interference type under the real interference type (the closer the value is to 100%, the more successful the identification is, indicating that the identification effect is better).

[0105] It can be seen from Figure 2 that the identification rates of the 7 interference types are all higher than 99.67% under the constructed complex electromagnetic environment, the interference identification method proposed in the present application performs excellently in identification accuracy, stability and multi-type interference distinguishing ability, and verifies the accurate identification ability of the method to radar active jamming under the complex electromagnetic condition.

[0106] In the training process of feature extraction of the method, the generation of training data is not limited to specific specific jamming noise ratio and other interference parameters, but a randomized interference identification training framework is proposed, random jamming noise ratio and other parameters are used to simulate a complex electromagnetic environment, and the problem of large difference in interference identification effect of different conditions of the traditional interference identification classification model is avoided, in the interference identification process, a radar interference signal classification and identification method based on multi-dimensional, multi-domain and multi-frame feature extraction is proposed, the characteristics of radar signals in each dimension and each transform domain are utilized, multi-dimensional and multi-domain feature information of multi-frame data is comprehensively extracted, a high-dimensional feature vector is synthesized, a distance-velocity dimension motion state mapping matrix is combined, and the type of active interference signal is effectively identified.

Claims

1. A radar interference signal classification and identification method based on multi-dimensional, multi-domain, and multi-frame feature extraction, characterized in that, The identification method includes the following steps: Step 1: Build a randomized interference recognition training framework and construct a recognition sample library covering a continuous parameter space; Step 2: Cross-domain joint extraction of multi-dimensional signal features to extract distinguishable features of interference from different radiation sources in multiple feature domains, forming a high-dimensional feature vector. Step 3: Monitor the radial distance change rate and echo Doppler frequency to construct a distance-velocity dimension motion state mapping matrix; Step 4: Fuse high-dimensional feature vectors and mapping matrices to train a classification and recognition model; Step 5: Output the radiation source interference type label.

2. The radar interference signal classification and identification method according to claim 1, characterized in that, The steps for building the randomized interference recognition training framework described in step 1 are as follows: The first step is to construct environmental parameters covering a continuous range: ； in, Indicates the signal-to-noise ratio. Indicates clutter intensity. Indicates signal bandwidth. Indicates the signal carrier frequency. This represents the signal pulse width. All environmental parameters follow a continuous interval distribution, with a value range of [value range missing]. Together, they constitute a continuous 5-dimensional environmental parameter space; The second step is to construct the interference parameters covering continuous intervals: ； in, Indicates the type of interference. Indicates the noise-to-interference ratio (NIR). Indicates the interference bandwidth. This represents the interference frequency. All interference parameters follow a continuous interval distribution, with a value range of [value range missing]. Together, they constitute a continuous space of 4-dimensional interference parameters.

3. The radar interference signal classification and identification method according to claim 2, characterized in that, The steps for constructing the recognition sample library covering the continuous parameter space described in step 1 are as follows: The first step is to select a combination of continuous environmental parameters and interference parameters based on a randomized interference recognition training framework and a preset step size. The second step is to generate the corresponding radiation source interference signal based on the selected parameter combination; The third step is to store the generated radiation source interference signal in accordance with its corresponding environmental parameters, interference parameters and interference type; The fourth step is to arrange the samples in order of parameter size to form a sample library for radar interference signal identification.

4. The radar interference signal classification and identification method according to claim 1, characterized in that, The multi-dimensional signal feature cross-domain joint extraction mentioned in step 2 refers to extracting distinguishable features of different radiation source interference in multiple feature domains such as time domain, frequency domain, and time-frequency domain through multi-dimensional feature space mapping and correlation analysis, forming a high-dimensional feature tensor representation.

5. The radar interference signal classification and identification method according to claim 1, characterized in that, The high-dimensional feature vector mentioned in step 2 is as follows: ； in, This indicates the sharp characteristics of the time-domain waveform. Indicates the skewness characteristics of the time-domain waveform. This indicates the sharp characteristics of the frequency domain waveform. Indicates the skewness characteristics of the frequency domain waveform. This represents the temporal envelope undulation characteristics, and the superscript T indicates the transpose operation.

6. The radar interference signal classification and identification method according to claim 1, characterized in that, The radial distance change rate mentioned in step 3 is as follows: ； in, and They represent the first Frame and the Target radial distance extracted from the frame. This is the set total number of monitoring frames.

7. The radar interference signal classification and identification method according to claim 6, characterized in that, The distance-velocity dimension motion state mapping matrix mentioned in step 3 is as follows: ； in, Indicates the first Frame and the The rate of change of frame speed and They represent the first Frame and the The target radial velocity extracted from the frame. This is the set total number of monitoring frames.

8. The radar interference signal classification and identification method according to claim 7, characterized in that, The fusion of high-dimensional feature vectors and mapping matrix described in step 4 is as follows: ； in, This represents the high-dimensional feature vector obtained by cross-domain joint extraction of multi-dimensional signal features. This represents the distance-velocity dimension motion state mapping matrix. and Denotes the fusion weight coefficient, and satisfies .

9. The radar interference signal classification and identification method according to claim 1, characterized in that, The steps for training the classification and recognition model in step 4 are as follows: The first step is to obtain the high-dimensional feature vector obtained by cross-domain joint extraction of multi-dimensional signal features, and the range-velocity motion state mapping matrix obtained based on multi-frame radar signal data; The second step is to construct a multi-scale span fusion perception classification model, and to fuse the high-dimensional feature vector and the motion state mapping matrix at the feature level according to a preset ratio to obtain fused features. The third step is to use the fused features as model input, perform layer-by-layer feature mapping on the fused features, and output the corresponding interference type prediction results. The fourth step involves calculating the loss function based on the difference between the model's predicted output and the actual interference type labels, and then training the classification and recognition model by iteratively updating the model parameters until the model converges. The loss function for calculating the difference is as follows: ； in, This represents the total number of feature scales involved in decision fusion in a multi-scale span fusion perceptual classification model. Indicates the first The loss weight coefficients corresponding to each scale Indicates the first The interference type discrimination vector output at each scale This represents the true interference type label vector for the corresponding sample. This represents a difference measurement function used to measure the deviation between the predicted results and the true labels.

10. The radar interference signal classification and identification method according to claim 9, characterized in that, The multi-scale span fusion perception classification model consists of a cascaded input layer, a hidden layer, and an output layer. The input layer is used to receive fused features. The fused features are formed by fusing high-dimensional feature vectors obtained by cross-domain joint extraction of multi-dimensional signal features with a range-velocity motion state mapping matrix constructed based on multi-frame radar signal data at the feature level according to a preset ratio. The fused features are input into the model in the form of a two-dimensional feature matrix, where each row corresponds to a sample and each column corresponds to a feature. Hidden layers are used to perform layer-by-layer feature mapping and multi-scale feature extraction on the input fused features; the hidden layer consists of three feature mapping sub-layers, each of which corresponds to a set of weight parameter matrices. and bias parameter vector The initial value is set to and , Indicates a uniform distribution; The output layer is used to classify and discriminate the multi-scale span features output by the third hidden layer. Its parameters include the classification weight parameter matrix and the classification bias parameter vector. The output results are used to represent the discrimination results of each radar interference type. During the model training phase, the sample input to the multi-scale span fusion perceptual classification model passes through the input layer, hidden layer, and output layer in sequence, resulting in the interference type discrimination vector of the output layer. Compare it with the actual interference type label vector Perform a difference metric to calculate the loss function. Then, gradient calculations are performed on the hidden layer model parameters, and the update step size is set according to the preset step size. Iterative updates are performed on the parameters of each layer to make the loss function... It gradually decreases as the number of iterations increases; The training termination condition is set as follows: when the difference between the loss functions of two adjacent iterations is less than the preset threshold of 0.001, or when the number of iterations reaches the preset upper limit of 100,000, training stops and the converged model parameters are output.

Images