Radar active jamming identification and classification method based on multi-channel TCN-BiGRU network fusion

Abstract

The invention belongs to the technical field of radars, and discloses a TCN-BiGRU fusion network-based radar active jamming signal identification method, which comprises the following steps of: establishing different radar active jamming signal models; performing wavelet packet decomposition on the radar active interference signal to obtain a wavelet packet decomposition coefficient; respectively establishing data sets for the real part, the imaginary part and the module value of the wavelet packet decomposition coefficient, and dividing the data sets into a training set and a verification set; establishing a three-channel TCN-BiGRU network model, and training to obtain a radar active jamming signal identification model; and inputting the test set and the verification set into a radar active interference signal identification model for signal identification and verification. According to the method, wavelet packet decomposition is adopted, a real part, an imaginary part and a module value of an interference signal are separated, and three groups of independent data sets are formed. A time convolution network TCN and a bidirectional gating loop unit BiGRU are connected in series, a three-channel network model is established, three groups of data are processed respectively, and outputs of the three channels are fused through weighted voting, so that the overall recognition accuracy of the model is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal recognition, and in particular relates to a radar active interference recognition and classification method based on a multi-channel TCN-BiGRU network fusion model. Technical Background

[0002] In the application of modern radar systems, accurate identification and processing of active jamming signals are particularly critical. Active jamming signals, such as electronic countermeasure signals, usually take a variety of complex forms, including suppression jamming and deceptive jamming, aiming to disrupt the normal function of the radar system. Interference identification methods based on feature extraction have achieved good results in specific scenarios, but such methods rely on manual screening of feature parameters, which is insufficient when dealing with interference in highly dynamic and complex environments. With the rapid development of deep learning technology, especially in the field of time series analysis and signal processing, deep learning-based methods have provided new solutions to this problem. In particular, the fusion model of temporal convolutional network (TCN) and gated recurrent unit (BiGRU), using the powerful time series data processing capability of TCN and the advantages of BiGRU in processing long-term dependencies, shows great potential for identifying radar jamming signals in complex interference environments.

[0003] Interference identification methods based on a single network have limitations in fully capturing the comprehensive characteristics of complex interference signals and in their generalization capabilities under unknown interference types or changing environments. These limitations lead to insufficient recognition accuracy in practical applications.

[0004] Therefore, it is an urgent problem to provide a multi-network fusion model to realize a radar interference signal recognition method with high recognition rate in complex environments. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and propose a radar active interference identification and classification method based on a multi-channel TCN-BiGRU network fusion model.

[0006] In order to achieve the above object, the present invention adopts the following technical methods to achieve it.

[0007] The radar active jamming signal recognition method based on the multi-channel TCN-BiGRU network fusion model includes the following steps:

[0008] Step 1, establishing different radar active jamming signal models;

[0009] Step 2, generating three different data sets of real part, imaginary part and modulus value of wavelet packet decomposition coefficients after wavelet packet decomposition of radar active interference signal;

[0010] Step 3, divide the data set into training set and test set in proportion;

[0011] Step 4, construct a multi-channel TCN-BiGRU network fusion model;

[0012] Step 5: The training set is input into the multi-channel TCN-BiGRU network fusion model to obtain the radar active jamming signal recognition model;

[0013] Step 6: Input the test set multi-channel TCN-BiGRU network fusion model and perform weighted fusion of the recognition results of the three channels to identify radar active interference signals.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] (1) Compared with the conventional active interference signal recognition method based on neural network, the radar active interference signal recognition method based on the multi-channel TCN-BiGRU network fusion model of the present invention does not need to perform a series of signal preprocessing such as filtering, denoising, and normalization on the active interference signal.

[0016] (2) The radar active interference signal identification method based on the multi-channel TCN-BiGRU network fusion model of the present invention can complete the training of the fusion network by only using the wavelet packet decomposition coefficient data of each active interference signal after wavelet packet decomposition, and the sample is simple and the data complexity is small.

[0017] (3) The radar active interference signal recognition method based on the multi-channel TCN-BiGRU network fusion model of the present invention constructs a multi-channel series network, TCN extracts spatiotemporal features, and BiGRU extracts long-term dependency features and context features. In addition, the multi-channel input design further enhances the model's ability to handle different types of interference signals, thereby improving the recognition accuracy and response speed of interference signals, especially showing excellent performance when dealing with a variety of complex and dynamically changing interference environments, ensuring an efficient and stable recognition process. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] Figure 1 This is a flow chart of the radar active jamming signal identification method based on the multi-channel TCN-BiGRU network fusion model;

[0020] Figure 2 It is a multi-channel TCN-BiGRU network fusion structure diagram;

[0021] Figure 3 is the confusion matrix of the recognition results;

[0022] Figure 4This is a comparison chart of the recognition accuracy of the multi-channel TCN-BiGRU network fusion and the traditional convolutional neural network model; DETAILED DESCRIPTION

[0023] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0024] like Figure 1 As shown, the radar active jamming signal recognition method based on the multi-channel TCN-BiGRU network fusion model includes the following steps:

[0025] Step 1, establishing different radar active jamming signal models;

[0026] Specifically, a space containing real echo signals, radar interference signals and additive Gaussian white noise signals is established. Ten different radar active interference signal models are generated, including suppression interference and deception interference, among which radar active deception interference includes dense forwarding interference, slicing interference, distance deception interference, speed deception interference, and distance-speed deception interference; active suppression interference includes noise frequency modulation interference, noise amplitude modulation interference, noise convolution interference, noise product interference and radio frequency noise interference. The interference signal model can be expressed as:

[0027] r(n)=s(n)+j(n)+y(n) (1)

[0028] Among them, s(n) represents the real echo signal, j(n) represents the interference signal, and y(n) represents the additive Gaussian white noise.

[0029] Step 2, generating a data set of radar active jamming signals;

[0030] Sub-step 2.1, perform wavelet packet decomposition on each radar interference signal, and the obtained wavelet packet decomposition coefficients are expressed as:

[0031]

[0032] Among them, d i，j，2m ,d i,j,2m+1 represents the wavelet packet coefficient of each node, g(k) represents a low-pass filter, and h(k) represents a high-pass filter.

[0033] Sub-step 2.2, by changing the relevant parameters of different radar active interference such as repetition interval, interference amplitude and frequency, a complete wavelet packet decomposition coefficient data set is generated, and a label is set for each type of interference. The bandwidth range of each type of interference signal is 1MHz~4MHz, the interval is 1, the interference amplitude range is 10~40, the interval is 10, and it is cycled 5 times. Each type of interference generates a total of 4×4×5 samples, and 10 types of interference generate a total of 4×4×5×10 interference signal data. Finally, a label is set for each type of interference according to the sequence number 1-10.

[0034] Step 3, divide the data set into training set and test set;

[0035] Sub-step 3.1, randomly extract each type of signal from the radar active interference signal wavelet packet decomposition coefficient data set, and combine them into a training set and a test set in a ratio of 8:2.

[0036] Step 4, construct a multi-channel TCN-BiGRU network fusion model;

[0037] Sub-step 4.1, first initialize the blank network structure, then establish the input layer and sequence folding layer, and add the established network structure to the blank network structure;

[0038] Sub-step 4.2, establish a temporal convolutional network (TCN) module, whose structure is as follows: the first convolution layer, the convolution layer uses dilated convolution, and the number of convolution kernels is 64. The second activation layer, the activation layer uses the Relu function as the activation function. The third residual connection layer, the residual block introduces skip connections, allowing part of the network input to directly "skip" some layers to be passed to deeper layers. This structure allows the network to learn the complex features of the data more deeply while maintaining a low training error. The fourth normalization layer is used to improve the model's feature extraction and learning capabilities on time series data;

[0039] Sub-step 4.3, build a gated recurrent unit network (BiGRU) module, the structure is as follows: the first BiGRU layer, used to process sequence data and capture temporal dependencies. The second fully connected layer, converts the BiGRU output into features suitable for classification. The third softmax classification layer, used to output probability distribution and perform multi-class classification;

[0040] Sub-step 4.4, connect the temporal convolutional network module and the gated recurrent unit network module in series to generate a TCN-BiGRU network fusion model;

[0041] Sub-step 4.5, construct a three-channel TCN-BiGRU network fusion model, fuse the output results of the three models using weighted voting, and establish a good fusion network model such as Figure 2 shown.

[0042] Step 5: Train the three-channel TCN-BiGRU network fusion model to obtain a radar active jamming signal recognition model;

[0043] Sub-step 5.1, shuffling the interference signals in the divided training set to obtain a shuffled training set;

[0044] In sub-step 5.2, set the learning rate of training to 0.001 and the maximum number of times the network model is trained to 500. Adopt the Adam optimizer to iteratively update the parameters of the fusion network, and improve the training process by adaptively adjusting the learning rate and using momentum to help the model converge faster and be more stable;

[0045] Sub-step 5.4, each group of interference signals (x 1 , x 2 ,...x 999 , x 1000 ) is input into the TCN-BiGRU network fusion model, firstly the local features in the time series are captured by the TCN layer, and then these features are sent to the BiGRU network for global feature extraction. The BiGRU contains the input x at this time i and the hidden state h left at the previous moment i-1 Through these two nodes, we can get two gate states: the gate r for controlling the reset and the gate z for controlling the update, which can be expressed mathematically as

[0046]

[0047] Where W r , W z and U r , U z is the input of the gate signal and the hidden state coefficient. h is reset by r i-1′ =h i-1 ⊙r.

[0048] Sub-step 5.5, using the input data x i 、Data after reset h i-1′ The tanh activation function converts the calculated data to between -1 and 1. The mathematical expression is as follows

[0049] h′=tanh(W h x i +U h h i-1′+ b h ) (4)

[0050] Sub-step 5.6, "update memory" through the update gate z in the range of 0 to 1, including forgetting function and memory function. The larger the gate signal z, the more "memory" there is, otherwise the more "forgotten" data there is.i-1 The unimportant information in the dimension is forgotten using the "forget gate" 1-z, expressed as 1-z⊙h i-1 . Use z⊙h′ to selectively “memorize” the current node, and the updated mathematical expression is as follows

[0051] h i =1-z⊙h i-1 +z⊙h (5)

[0052] In the formula, h i Represents the current hidden state. The weight corresponding to the current input h′ makes up for the forgotten dimensional information weight, making the network in a "constant state", so z and 1-z are linked.

[0053] Step 6: Input the test set to identify the radar active jamming signal;

[0054] Sub-step 6.1, shuffle the test set, input the shuffled test set into the trained three-channel TCN-BiGRU network fusion model for recognition, and obtain the classification result.

[0055] The effect of the present invention is further described below in conjunction with simulation experiments.

[0056] 1. Simulation conditions

[0057] The radar active jammer used in the simulation experiment of the present invention includes 10 types of radar jammers, namely dense forwarding jammer, slicing jammer, distance deception jammer, speed deception jammer, distance-speed deception jammer, noise frequency modulation jammer, noise amplitude modulation jammer, noise convolution jammer, noise product jammer and radio frequency noise jammer. The pulse repetition period of each jammer is 1000μs, the number of pulses is fixed to 1, the sampling frequency is fixed to 2MHz, the bandwidth range of each jammer is 1MHz to 4MHz, and the interference amplitude range is 10 to 40.

[0058] 2. Simulation content and result analysis

[0059] The training set and test set used in the simulation experiment of the present invention are 10 types of radar active interference signals extracted from the radar active interference signal wavelet packet decomposition coefficient data set generated in step 2 of the present invention. According to the ratio of 8:2 between the training set and the test set of each interference signal, the training set has a total of 1600 samples, 160 samples for each type of interference training, and the test set has a total of 400 samples, 40 samples for each type of interference test.

[0060] The 10 types of radar jammers are: dense forwarding jammer, slicing jammer, distance deception jammer, speed deception jammer, distance-speed deception jammer, noise FM jammer, noise AM jammer, noise convolution jammer, noise product jammer and RF noise jammer.

[0061] Use the multi-channel TCN-BiGRU network fusion model trained with the training set, input the test set at an interference-to-noise ratio of 10dB to identify interference signals, and output the classification results.

[0062] Identify the confusion matrix and accuracy curve as follows Figure 3 and Figure 4 shown.

[0063] Figure 3 The labels 1-10 in the figure correspond to 10 types of radar interference signals. The labels from 1 to 10 are: dense forwarding interference, slicing interference, distance deception interference, speed deception interference, distance-speed deception interference, noise frequency modulation interference, noise amplitude modulation interference, noise convolution interference, noise product interference and radio frequency noise interference.

[0064] from Figure 3 It can be seen that the recognition rate of 8 types of radar interference signals reached 100%, and the overall recognition rate reached 99.5%, which further shows that the fusion network has a high recognition rate for radar active interference signals.

[0065] from Figure 4 It can be seen that the multi-channel TCN-BiGRU network fusion model proposed in the present invention has a higher recognition rate than the traditional convolutional neural network interference recognition method.

[0066] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A radar active jammer identification and classification method based on multi-channel TCN-BiGRU network fusion, comprising the following steps: Step 1, establishing different radar active jamming signal models; Step 2, generating three different data sets of real part, imaginary part and modulus value of wavelet packet decomposition coefficients after wavelet packet decomposition of radar active interference signal; Step 3, divide the data set into training set and test set in proportion; Step 4, construct a multi-channel TCN-BiGRU network fusion model; Step 5, input the training set into the multi-channel TCN-BiGRU network fusion model to obtain the radar active interference signal recognition model; Step 6: Input the test set multi-channel TCN-BiGRU network fusion model and perform weighted fusion of the recognition results of the three channels to identify radar active interference signals.

2. The radar active jamming signal identification method based on the multi-channel TCN-BiGRU network fusion model according to claim 1 is characterized in that: Step 1 contains the following sub-steps: Sub-step 1.1, generate 10 different radar active jamming signal models, including suppression jamming and deception jamming, where radar active deception jamming includes dense forwarding jamming, slicing jamming, distance deception jamming, speed deception jamming, and distance-speed deception jamming; active suppression jamming includes noise frequency modulation jamming, noise amplitude modulation jamming, noise convolution jamming, noise product jamming, and radio frequency noise jamming. The established jamming signal model includes real echo signal, radar jamming signal, and additive white Gaussian noise signal. The jamming signal model can be expressed as: r(n)=s(n)+j(n)+y(n) Among them, s(n) represents the real echo signal, j(n) represents the interference signal, and y(n) represents the additive Gaussian white noise.

3. The radar active jamming signal identification method based on the multi-channel TCN-BiGRU network fusion model according to claim 2 is characterized in that: Step 2 contains the following sub-steps: Sub-step 2.1, by changing the relevant parameters of different radar interferences such as repetition interval, interference amplitude and frequency, each radar interference signal is decomposed by wavelet packet to obtain the wavelet packet decomposition coefficients Sub-step 2.2, decompose the wavelet packet decomposition coefficients, extract the real part, imaginary part and modulus value respectively, form three data sets, and set a label for each type of interference.

4. The radar active jamming signal identification method based on the multi-channel TCN-BiGRU network fusion model according to claim 3 is characterized in that: Step 3 contains the following sub-steps: Sub-step 3.1, randomly extract each type of signal from the radar interference signal wavelet packet decomposition coefficient data set, and combine them into a training set and a test set in a ratio of 8:

2.

5. The radar active jamming signal identification method based on the multi-channel TCN-BiGRU network fusion model according to claim 4 is characterized in that: Step 4 contains the following sub-steps: Sub-step 4.1, first initialize the blank network structure, then establish the input layer and sequence folding layer, and add the established network structure to the blank network structure; Sub-step 4.2, establish a temporal convolutional network (TCN) module, whose structure is as follows: the first convolution layer, the convolution layer uses dilated convolution, and the number of convolution kernels is 64. The second activation layer, the activation layer uses the Relu function as the activation function. The third residual connection layer, the residual block introduces skip connections. This structure enables the network to learn the complex features of the data more deeply while maintaining a low training error. The fourth normalization layer is used to improve the model's feature extraction and learning capabilities on time series data; Sub-step 4.3, build a gated recurrent unit network (BiGRU) module, the structure is as follows: the first BiGRU layer, used to process sequence data and capture temporal dependencies. The second fully connected layer, converts the BiGRU output into features suitable for classification. The third softmax classification layer, used to output probability distribution and perform multi-class classification; In sub-step 4.4, the temporal convolutional network module and the gated recurrent unit network module are connected in series to generate a TCN-BiGRU network fusion model. Sub-step 4.5, construct a three-channel TCN-BiGRU network model and fuse the output results of the three models using weighted voting.

6. The radar active jamming signal identification method based on the multi-channel TCN-BiGRU network fusion model according to claim 5 is characterized in that: Step 5 contains the following sub-steps: Sub-step 5.1, shuffling the interference signals in the divided training set to obtain a shuffled training set; In sub-step 5.2, set the learning rate of training to 0.001 and the maximum number of times the network model is trained to 500. Adopt the Adam optimizer to iteratively update the parameters of the fusion network, and improve the training process by adaptively adjusting the learning rate and using momentum to help the model converge faster and more stably; Sub-step 5.3, each group of interference signals (x1, x2, ...x 999 , x 1000 ) is input into the TCN-BiGRU network fusion model, firstly the local features in the time series are captured by the TCN layer, and then these features are sent to the BiGRU network for global feature extraction. The BiGRU contains the input x at this time i and the hidden state h left at the previous moment i-1 Through these two nodes, two gate states can be obtained: the gate r that controls the reset and the gate z that controls the update, which is mathematically expressed as r=σ(W r x i +U r h i-1 +b r ) z=σ(W z x i +U z h i-1 +b z ) Where W r , W z and U r , U z are the input and hidden state coefficients of the gating signal respectively. Sub-step 5.4, reset r to get h i-1′ =h i-1 ⊙r. Using input data x i 、Data after reset h i-1′ The tanh activation function converts the calculated data to between -1 and 1. The mathematical expression is as follows h′=tanh(W h x i +U h h i-1′ +b h ) Sub-step 5.5, "update memory" is performed through the update gate z in the range of 0 to 1, including forgetting function and memory function. The larger the gate signal z, the more "memory" there is, otherwise the more "forgotten" data there is. i-1 The unimportant information in the dimension is forgotten using the "forget gate" 1-z, expressed as 1-z⊙h i-1 . Use z⊙h′ to selectively "memorize" the current node, and the updated mathematical expression is as follows In the formula, h i Represents the current hidden state. The weight corresponding to the current input h′ makes up for the forgotten dimensional information weight, making the network in a "constant state", so z and 1-z are linked.

7. The radar active jamming signal identification method based on the multi-channel TCN-BiGRU network fusion model according to claim 6 is characterized in that: Step 6 contains the following sub-steps: Sub-step 6.1, shuffle the test set, and input the shuffled test set into the trained three-channel TCN-BiGRU network fusion model for recognition. In sub-step 6.2, the recognition results of the three-channel TCN-BiGRU network fusion model are weighted and output in a certain proportion to obtain the final recognition result.

Images