A passive jamming identification method based on HRRP and SSM-Transformer

Abstract

The application discloses a passive jamming identification method based on HRRP and SSM-Transformer; the application firstly acquires an HRRP sequence through pulse compression or calculation of RCS and inverse fast Fourier transform (IFFT), then performs two-dimensional reconstruction and position coding on the HRRP sequence, and finally inputs the HRRP sequence into an SSM-Transformer network to complete training and prediction. The SSM module can extract long-term dependence relationship information between sequences, and the multi-head attention mechanism of the Transformer encoder can extract deep features in the sequence. The application can improve the identification accuracy to a certain extent through the expansion of the network structure, and has better identification and generalization performance.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar passive jamming recognition, and particularly relates to a passive jamming recognition method based on HRRP and SSM-Transformer. BACKGROUND

[0002] Jamming recognition is an important part of the field of radar technology in modern battlefield, and is the premise of implementing radar anti-jamming measures. The passive jamming represented by corner reflector array and chaff cloud scatters radar transmitted signals to create multiple false targets, which has an inducing and deceptive effect on the radar, and it is difficult to effectively identify passive jamming and ships in a mixed scene.

[0003] The passive jamming recognition method based on high resolution range profile (HRRP) mainly uses a machine learning classifier to classify according to the HRRP features reflecting the target geometric structure and length information extracted. With the development of artificial intelligence technology, the HRRP recognition based on deep learning has become a research hotspot. Deep neural networks can automatically extract more deep features, and have better recognition performance, mainly including RNN and CNN. However, with the increase of HRRP sequence length and sample size, the traditional RNN faces the problems of gradient disappearance and gradient explosion, and the traditional CNN tends to extract local features, which are difficult to handle long-term dependencies, resulting in insufficient feature extraction. SUMMARY

[0004] In view of the above problems and inspired by the significant achievements of the Transformer in sequence modeling tasks such as natural language processing, the application proposes a passive jamming recognition method based on SSM-Transformer, which can fully extract time sequence features and improve recognition performance.

[0005] The passive jamming recognition method based on SSM-Transformer proposed by the application comprises the following steps:

[0006] Step S1, constructing a data set based on HRRP sequences of chaff cloud, corner reflector array and ship echo:

[0007] Step S1.1, performing pulse compression operation based on chaff cloud echo to obtain HRRP sequence;

[0008] Step S1.2, calculating RCS based on corner reflector array and ship echo through physical optical algorithm, and performing IFFT conversion to obtain HRRP sequence;

[0009] Step S2: Perform two-dimensional reconstruction of the HRRP sequence, divide the HRRP into multiple row vectors, and form a sequence vector matrix X0 to reduce the distance between different sampling points that are far apart.

[0010] Step S3: Based on position encoding, the order information between row and column vectors in X0 is represented so that each row and column vector has a sequential relationship.

[0011] Step S4: Construct the SSM module to extract the dependency information between each segment of the sequence;

[0012] Step S5: Use the Transformer encoder to extract features to obtain hidden layer representations for classification.

[0013] Furthermore, the dataset construction described in step S1 is based on the scattering center model. HRRP can be viewed as the superposition of the target scattering centers distributed across multiple range cells on the radar line of sight. Each scattering point corresponds to a specific geometric configuration, from which a large amount of identification information associated with the target structure can be extracted. The radar HRRP is a one-dimensional sequence, where each element represents the scattering amplitude of a range cell, which can be expressed as:

[0014] x = {|x1|,|x2|,|x3|,…,|x} n |,…,|x N |}, n=1,2,…,N (1)

[0015]

[0016] Where |·| represents modulo, and N represents the number of distance units in HRRP, i.e., the sequence length. S n f represents the number of scattering points within the nth range cell. c Indicates the radar carrier frequency. and Let represent the scattering intensity and echo delay of the i-th scattering point within the n-th range cell, respectively. The HRRP represents the incident elevation and azimuth angle of the radar relative to the scatterer, and it is sensitive to angle. As can be seen from equation (2), each element of the HRRP represents the scattering intensity at different scattering centers of the scatterer to the corresponding radial distance of the radar, reflecting the structural information of the scatterer and having a time-dependent relationship, thus providing the possibility for identifying ships and passive interference.

[0017] Furthermore, the two-dimensional reconstruction described in step S2 is for extracting long-term dependencies. The HRRP is segmented into multiple row vector inputs, thus reducing the distance between different sampling points that are far apart. The sequence vector matrix X0 can be represented as:

[0018]

[0019] wherein x 1×N is HRRP, D1 is the total number of divided segments, and each segment has a length of D2=N / D1.

[0020] Further, the position encoding in step S3 can represent the order information between row and column vectors in the two-dimensional reconstructed vector matrix X0. The present application encodes the position information in the form of cosine to ensure the uniqueness of the encoding, and the calculation formula of the position encoding PE is:

[0021]

[0022] wherein t=1, 2, …, D1 represents the position order, and 2i and 2i+1 represent the even dimension and odd dimension in 1 to D2 respectively. The sequence vector matrix X0 and its position encoding are added to obtain the input matrix X.

[0023] Further, the SSM in step S4 is a model representing the relationship between the state at a certain moment and the input and the state at the previous moment, and the SSM module is constructed before the multi-head attention layer to extract the dependency information between the front and rear sub-sequences, so that the attention mechanism can better assign weights to the target segment. The expression of SSM is as follows:

[0024]

[0025] wherein x k is the input vector, y k is the output vector, A, B, and C are system matrices, X is the input matrix, K is the convolution kernel, and y is the output matrix. The matrix A is related to the state vector, contains the information of the previous state to construct a new state, and also determines how these information is transmitted in time sequence. In order to better capture the information of the sub-sequences with shorter distance and attenuate the information of the sub-sequences with longer distance, the matrix A is created by using High-order Polynomial Projection Operators (HiPPO), and the HiPPO square matrix can be expressed as:

[0026]

[0027] wherein n is the row sequence number of the matrix, and k is the column sequence number of the matrix. The convolution kernel K is used to perform convolution with the input to obtain the time sequence relationship information between each row vector of the data, which is beneficial to extract more rich time sequence relationship features after being input into the Transformer encoder.

[0028] Further, the Transformer encoder in step S5 mainly includes a multi-head attention layer, a residual connection and layer normalization, and a feedforward fully connected layer. The multi-head attention layer in the Transformer encoder is used for feature extraction to obtain time sequence features for training and testing, so as to complete passive jamming recognition. The HRRP samples are divided into a training set and a test set in a ratio of 8:2. During training, the input sample batch is set to 8, the training round is set to 200 rounds, and the learning rate is set to 0.001. The multi-head attention layer is a structure for fusing multiple single self-attention extracted information. The input sequence is linearly transformed by being multiplied by different weight coefficients W Q 、W K and W V to obtain three input matrices of the attention mechanism, namely a query matrix Q, a key matrix K and a value matrix V. Based on Q, the correlation between Q and all key matrices K is calculated by the inner product of Q and all key matrices K to obtain a weight distribution. The greater the correlation is, the greater the weight of the corresponding value matrix V is, so as to represent the importance of different element information. For the input sequence x, the output of the attention mechanism can be represented as:

[0029]

[0030] wherein, (·) T represents matrix transposition, QK T represents the dot product of matrix Q and matrix K T , d k represents the dimension size of matrix K. The formula first calculates the similarity of Q and K, then performs scaling and softmax normalization on the similarity, and finally performs weighted summation on the V vector by using the normalized weight to obtain the output of the attention.

[0031] The residual connection and the layer normalization layer use the residual connection to improve the gradient disappearance problem in the back propagation process of the deep network, and use the layer normalization to transform the data into a standard distribution with a mean of 0 and a variance of 1.

[0032] The feedforward fully connected layer layer performs nonlinear transformation on the multi-head attention output, uses the ReLU function as the activation function, and further plays a role in feature extraction and increasing nonlinearity.

[0033] Advantages:

[0034] 1. Automatic feature extraction: the method based on the SSM-Transformer of the present application effectively reduces the dependence on manual feature extraction, automatically represents the time sequence dependence information of the HRRP sequence through the SSM module, and improves the efficiency and accuracy of feature extraction.

[0035] 2. Deep temporal correlation feature extraction: By utilizing the multi-head attention mechanism in the Transformer encoder, the invention can deeply mine the deep temporal correlation features in the sequence, thereby significantly improving the performance of passive jamming recognition.

[0036] 3. Improve recognition accuracy: By comprehensively utilizing temporal dependency information and deep feature extraction through the SSM module and the Transformer encoder, the invention achieves higher recognition accuracy in the passive jamming recognition task, providing more reliable technical support for related field applications. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 is the overall framework diagram of the SSM-Transformer-based passive jamming recognition method.

[0038] Figure 2 is the simulated HRRP of a chaff cloud, a corner reflector array, and a ship.

[0039] Figure 3 is the structure of the self-attention mechanism module.

[0040] Figure 4 is the comparison chart of recognition accuracy performance of different network structures.

[0041] Figure 5 is the recognition accuracy chart of datasets with different interference parameter distributions. DETAILED DESCRIPTION

[0042] To make the purpose, technical solutions, and advantages of the invention clearer and more apparent, the following will clearly and completely describe the invention by combining the embodiments of the present application and the accompanying drawings.

[0043] The detailed process of the method described in this embodiment is as follows:

[0044] Figure 1 is the overall framework diagram of the SSM-Transformer-based passive jamming recognition method. The overall process is divided into three parts: dataset construction, preprocessing, and model construction. The preprocessing is further divided into two-dimensional reconstruction and position encoding, and the model construction is divided into the SSM module and the Transformer encoder:

[0045] Step S1, based on the HRRP sequence obtained from the chaff cloud, corner reflector array, and ship echo to construct the dataset:

[0046] The data sets of chaff cloud jamming, corner reflector array jamming and ship targets are constructed by simulation. For chaff jamming, the kinetic model of single chaff is established first, and the motion trajectory is simulated according to the parameters such as wind speed, initial velocity and attitude angle distribution, the spatial distribution of multiple chaffs is generated, the chaff cloud jamming echo is superimposed, and the HRRP is obtained by pulse compression. For corner reflector array jamming, an eight-corner reflector and a twenty-corner reflector with a side length of 1m are used to construct linear array, square array and circular array configurations, and different spacing parameters are set. The RCS is calculated by physical optics algorithm, and the equivalent HRRP is obtained by IFFT conversion. The size of the ship is 80m*10m*15m, and the method is the same as that of the corner reflector array. The pitch angle range of the radar relative to the target is set to 30° to 60°, the azimuth angle range is set to 20° to 145°, and the HRRP sampling point number is 200. The normalized HRRP samples under different model parameters and angles are shown in FIG. 1. Figure 2

[0047] Step S2, the HRRP sequence is two-dimensionally reconstructed, the HRRP is segmented to form multiple row vectors, and a sequence vector matrix X0 is composed, so that the distance between different sampling points far away from each other is reduced. The sequence vector matrix X0 can be expressed as:

[0048]

[0049] wherein x 1×N is the HRRP, D1 is the total number of divisions, and the length of each segment is D2=N / D1.

[0050] Step S3, the order information between the row and column vectors in X0 is represented based on position encoding, so that each row and column vector has a front and back relationship. The present application encodes the position information in the form of sine and cosine to ensure the uniqueness of the encoding, and the calculation formula of the position encoding PE is:

[0051]

[0052] wherein t=1, 2, …, D1 represents the position order, and 2i and 2i+1 represent the even dimension and the odd dimension in 1 to D2 respectively. The sequence vector matrix X0 and its position encoding are added to obtain the input matrix X.

[0053] Step S4, an SSM module is constructed to extract the front and back dependency information of each segment subsequence:

[0054] SSM is a model representing the relationship between the state at a certain time and the input and the state at the previous time. The SSM module is constructed before the multi-head attention layer to extract the front and back dependency information of each segment subsequence, so that the attention mechanism can better weight the target segment. The expression of SSM is as follows:

[0055] ​

[0056] Where, x k Let y be the input vector. k Let A, B, and C be the system matrices, X be the input matrix, K be the convolution kernel, and y be the output matrix. Matrix A is related to the state vector, containing information from previous states to construct new states, and it also determines how this information is propagated in time. To better capture information from nearby subsequences and attenuate information from distant subsequences, a high-order polynomial projection operator (HiPPO) is used to create matrix A. The HiPPO square matrix can be represented as:

[0057]

[0058] Where n is the row number of the matrix and k is the column number of the matrix. Convolving the input with kernel K yields the temporal relationship information between each row vector of the data. Feeding this information into the Transformer encoder helps extract richer temporal relationship features.

[0059] Step S5: Temporal features are extracted using the multi-head attention layer in the Transformer encoder for training and testing, completing the passive interference recognition. HRRP samples are divided into training and testing sets in an 8:2 ratio. During training, the input sample batch size is set to 8, the training epochs are set to 200, and the learning rate is 0.001.

[0060] A multi-head attention layer is a structure that fuses information extracted from multiple individual self-attention branches, as shown in the diagram below. Figure 3 As shown in the figure. The input data matrix in the figure is processed by different weighting coefficients W. Q W K and W V Multiplication and linear transformation yield the three input matrices of the attention mechanism: query matrix Q, key matrix K, and value matrix V. Based on Q, the correlation between Q and all key matrices K is calculated using their inner products, resulting in a weight distribution. Higher correlation indicates higher weights in the corresponding value matrix V, thus representing the importance of different elements. For an input sequence x, the output of the attention mechanism can be expressed as:

[0061]

[0062] in,(·) T QK represents the matrix transpose. T Representing matrices Q and K T Perform the dot product, d kThe dimension size of the matrix K is represented. The formula first calculates the similarity of Q and K, then scales and softmax normalizes the similarity, and finally weights the V vector by the normalized weight to obtain the output of the attention. Finally, the recognition result is output after the full connection layer and the softmax function.

[0063] Without loss of generality, the recognition performance is evaluated using the accuracy index.

[0064] Figure 4 The recognition performance of the algorithm under different network structures is shown, that is, the recognition accuracy under different numbers of Transformer encoders and the number of attention heads of the multi-head attention mechanism in the encoder. When 8 multi-head attention heads and different numbers of encoder layers are used, the change of recognition accuracy with JNR is as shown in Figure 4 (a). When 6-layer encoders and different numbers of multi-head attention heads are used, the change of recognition accuracy with JNR is as shown in Figure 4 (b).

[0065] As can be seen from Figure 3 , the recognition accuracy generally increases with the increase of JNR, and when JNR is greater than 20 dB, the recognition accuracy is greater than 90%. Increasing the number of encoder layers and attention heads can improve the recognition accuracy to a certain extent, and when 10-layer encoders and 8 attention heads or 6-layer encoders and 16 attention heads are used, and JNR is greater than 20 dB, the recognition accuracy is greater than 95%.

[0066] In order to verify the superiority of the proposed method compared with other deep learning-based recognition methods, the Transformer without adding the SMM module, ResNet18 and LSTM are compared. Among them, the Transformer network structure uses 10-layer encoders and 16 attention heads. In order to verify the strong generalization ability of the proposed method compared with other deep learning-based recognition methods, four data sets with different interference parameter distributions are simulated and constructed, and the change of interference recognition performance is analyzed. The interference parameter types and parameter values in different data sets are shown in Table 1.

[0067] Table 1

[0068]

[0069] The neural network is trained using the same distribution as the training set of data set a, and then tested using the four data sets respectively. For different recognition methods, the recognition accuracy and the decline amplitude of the recognition rate compared with data set a are counted, and the results are shown in Figure 5 .

[0070] As can be seen from Figure 5It can be seen that the SSM-Transformer method has a recognition accuracy of 98% on the same type of data set, which is higher than that of the comparative method. In terms of recognition generalization ability, the maximum decline of the recognition accuracy of the SSM-Transformer method on different types of data sets is 9.3%, which is lower than that of the comparative method. It is proved that the SSM-Transformer method has good recognition accuracy and generalization ability.

[0071] In summary, the SSM-Transformer-based passive jamming recognition method proposed in this paper can fully extract the time sequence relationship and improve the recognition performance.

[0072] The above description shows and describes several preferred embodiments of the present application, but as previously described, it should be understood that the present application is not limited to the forms disclosed herein, should not be considered as excluding other embodiments, and can be used in various other combinations, modifications and environments, and can be modified within the scope of the inventive concept described herein by the above teaching or related technical or knowledge. Any modification and change made by those skilled in the art without departing from the spirit and scope of the present application shall be within the protection scope of the claims of the present application.

Claims

1. A passive jamming identification method based on HRRP and SSM-Transformer, characterized in that, The method comprises the following steps: Step S1, constructing a dataset based on a chaff cloud, an array of corner reflectors, and a ship echo to obtain an HRRP sequence; Step S2, two-dimensional reconstruction is performed on the HRRP sequence, the HRRP is segmented to form multiple row vectors, and a sequence vector matrix is composed different sampling points far apart are reduced in distance; Step S3, encoding the order information between the row and column vectors based on the position The order information between the row and column vectors is represented, so that each row and column vector has a front-back relationship. Step S4, constructing an SSM module to extract the dependence information before and after each segment subsequence, wherein the expression of the SSM is as follows: ； where, is the input vector, is the output vector, , , are system matrices, is the input matrix, is the convolution kernel, is the output matrix; matrix is related to the state vector, which contains the information of previous states to build a new state, and it also determines how this information is transmitted in time series; in order to better capture the information of the subsequence with a short distance and attenuate the information with a long distance, a high-order polynomial projection operator (HiPPO) is used to create the matrix , and the HiPPO square matrix can be expressed as: ； wherein, is a matrix row number, is a matrix column number; a convolution kernel is used Convolution with the input obtains the timing relationship information between each row vector of the data, which is beneficial to extract more rich timing relationship features after being sent into the Transformer encoder. Step S5, using a Transformer encoder to perform feature extraction to obtain a hidden layer expression for classification.

2. The HRRP and SSM-Transformer based passive jamming identification method of claim 1, wherein, The specific method of step S1 is as follows: A dataset of chaff cloud jamming, corner reflector array jamming, and a ship target is constructed through simulation; for chaff jamming, a single chaff is first modeled, and a motion trajectory is simulated according to parameters such as wind speed, initial speed, and attitude angle distribution to generate a spatial distribution of multiple chaffs, and the chaff cloud jamming echo is obtained by superposition, and HRRP is obtained through pulse compression; for corner reflector array jamming, an eight-corner reflector and a twenty-corner reflector with a side length of 1 m are used to construct linear array, square array, and circular array configurations, and different spacing parameters are set, the RCS is calculated through physical optics algorithm, and IFFT conversion is performed to obtain the equivalent HRRP; the ship size is 80 m x 10 m x 15 m, and the method is the same as that of the corner reflector array; the pitch angle range of the radar relative to the target is set to 30° to 60°, the azimuth angle range is set to 20° to 145°, and the HRRP sampling point number is 200.

3. The HRRP and SSM-Transformer based passive jamming identification method of claim 1, wherein, The specific method of step S2 is as follows: The HRRP sequence is two-dimensionally reconstructed, the HRRP is segmented to form multiple row vectors, and the sequence vector matrix is composed The different sampling points with a relatively far distance are reduced in distance; the sequence vector matrix may be expressed as: ； wherein, is HRRP, is the total number of segments, each of length .

4. The HRRP and SSM-Transformer based passive jamming identification method of claim 1, wherein, In step S3, the position information is encoded in the form of sine and cosine to ensure the uniqueness of the encoding.

5. The HRRP and SSM-Transformer based passive jamming identification method of claim 1, wherein, Position encoding in step S3 The calculation formula is: ； wherein, denotes the position order, and denotes the even and odd dimensions, respectively, in ; adding the sequence vector matrix and its position encoding results in the input matrix .

6. The HRRP and SSM-Transformer based passive jamming identification method of claim 1, wherein, The Transformer encoder in step S5 mainly includes a multi-head attention layer, a residual connection and layer normalization, and a feedforward fully connected layer.

7. The HRRP and SSM-Transformer based passive jamming identification method of claim 1, wherein, In step S5, the multi-head attention layer in the Transformer encoder is used for feature extraction to obtain time sequence features for training and testing, and passive jamming recognition is completed; the HRRP samples are divided into a training set and a test set in a ratio of 8:2; during training, the input sample batch is set to 8, the training round is set to 200 rounds, and the learning rate is set to 0.

001.

8. The HRRP and SSM-Transformer based passive jamming identification method of claim 1, wherein, In step S5, for the input sequence The output of the attention mechanism is represented as: ； wherein, denotes matrix transpose, denotes matrix denotes matrix performs dot product, denotes matrix dimension size; the formula first calculates the similarity of , then scales and performs softmax normalization on the similarity, and finally performs weighted summation on vectors using the normalized weights to obtain the output of attention; finally, the recognition result is output through a fully connected layer and a softmax function.

Images