Passive interference identification method based on HRRP and SSM-Transformer

Abstract

The invention discloses a passive interference identification method based on an HRRP (High Rate Redundancy Protocol) and an SSM-Transform. According to the method, firstly, an HRRP sequence is obtained through pulse compression or RCS calculation and IFFT, then two-dimensional reconstruction and position coding are conducted on the HRRP sequence, and finally the HRRP sequence is input into an SSM-Transfomer network to complete training and prediction. Wherein the SSM module can extract long-term dependency relationship information between sequences, and the multi-head attention mechanism of the Transform encoder can extract deep features in the sequences. According to the method, the recognition accuracy can be improved to a certain extent by expanding the network structure, and better recognition and generalization performance is achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of radar passive interference identification, and in particular relates to a passive interference identification method based on HRRP and SSM-Transformer. Background Art

[0002] Interference identification is an important part of radar technology in modern battlefields and a prerequisite for the implementation of radar anti-interference measures. Passive interference represented by corner reflector arrays and chaff clouds creates multiple false targets by scattering radar transmission signals, which induces and deceives radars. It is difficult to effectively identify passive interference and ships in mixed scenarios.

[0003] The passive interference recognition method based on high resolution range profile (HRRP) mainly uses machine learning classifiers to classify according to the HRRP features extracted to reflect the target geometry and length information. With the development of artificial intelligence technology, HRRP recognition based on deep learning has become a hot research direction. 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, traditional RNN faces the problems of gradient vanishing and gradient exploding, and traditional CNN tends to extract local features, and it is difficult to handle long-term dependencies, resulting in insufficient feature extraction. Summary of the invention

[0004] In response to the above problems and inspired by the remarkable achievements of Transformer in sequence modeling tasks such as natural language processing, this paper proposes a passive interference recognition method based on SSM-Transformer, which can fully extract timing features and improve recognition performance.

[0005] The present invention proposes a passive interference identification method based on SSM-Transformer, comprising the following steps:

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

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

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

[0009] Step S2, reconstructing the HRRP sequence in two dimensions, segmenting the HRRP into multiple row vectors, and forming a sequence vector matrix X0, so that the distance between different sampling points that are far apart is reduced;

[0010] Step S3, characterizing the order information between the row and column vectors in X0 based on the position coding, so that each row and column vector has a front-and-back relationship;

[0011] Step S4, constructing an SSM module to extract the front-end dependency information of each subsequence;

[0012] Step S5, using the Transformer encoder to extract features and obtain hidden layer expressions for classification.

[0013] Furthermore, the data set constructed in step S1 is based on the scattering center model. HRRP can be regarded as the projection superposition of the target scattering centers distributed in multiple range units 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. Radar HRRP is a one-dimensional sequence, and each element in the sequence represents a scattering amplitude of a range unit, which can be expressed as:

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

[0015]

[0016] Here, |·| represents the modulus, and N represents the number of distance units in HRRP, i.e., the sequence length. n represents the number of scattering points within the nth distance unit, f c represents the radar carrier frequency, and represent the scattering intensity and echo delay of the ith scattering point in the nth range unit, respectively, and It represents the incident pitch and azimuth angle of the radar relative to the scatterer. HRRP is sensitive to angles. As can be seen from formula (2), each element of HRRP represents the scattering intensity at the corresponding radial distance from the different scattering centers of the scatterer to 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 in step S2 is to extract long-term dependencies, segment the HRRP into multiple row vector inputs, and reduce the distance between different sampling points that are far apart. The sequence vector matrix X0 can be expressed as:

[0018]

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

[0020] Furthermore, the position code in step S3 can characterize the order information between the row and column vectors in the two-dimensional reconstructed vector matrix X0. The present invention uses the sine and cosine form to encode the position information to ensure the uniqueness of the code, and the calculation formula of the position code PE is:

[0021]

[0022] Where t=1,2,…,D1 represents the position order, 2i and 2i+1 represent the even and odd dimensions from 1 to D2, respectively. The input matrix X is obtained by adding the sequence vector matrix X0 and its position encoding.

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

[0024]

[0025] Among them, x k is the input vector, y k is the output vector, A, B, C are all system matrices, X is the input matrix, K is the convolution kernel, and y is the output matrix. Matrix A is related to the state vector, contains the information of the previous state to construct the new state, and it also determines how this information is transmitted in the time series. In order to better capture the information of subsequences with closer distances and attenuate the information with farther distances, high-order polynomial projection operators (HiPPO) are used to create matrix A. The HiPPO square matrix can be expressed as:

[0026]

[0027] Among them, n is the matrix row number, and k is the matrix column number. The convolution kernel K is used to convolve with the input to obtain the temporal relationship information between each row vector of the data, which is sent to the Transformer encoder to help extract richer temporal relationship features.

[0028] Furthermore, the Transformer encoder described in step S5 mainly includes a multi-head attention layer, residual connection and layer normalization, and a feedforward fully connected layer. The multi-head attention layer in the Transformer encoder is used to extract features to obtain time series features for training and testing to complete passive interference identification; the HRRP samples are divided into training sets and test sets in a ratio of 8:2; during training, the input sample batch is set to 8, the training rounds are 200 rounds, and the learning rate is 0.001. The multi-head attention layer is a structure that fuses the self-attention extraction information of multiple single branches. The input sequence is combined with different weight coefficients W Q , W K and W V Multiply them for linear transformation to get the three input matrices of the attention mechanism: query matrix Q (query), key matrix K (key) and value matrix V (value). Based on Q, the correlation between Q and all key matrices K is calculated by the inner product between the two to get a weight distribution. The greater the correlation, the greater the weight of the corresponding value matrix V, which indicates the importance of different element information. For the input sequence x, the output of the attention mechanism can be expressed as:

[0029]

[0030] in,(·) T represents matrix transpose, QK T Represents matrix Q and matrix K T Perform the dot product, d k Represents the dimension size of the matrix K. The formula first calculates the similarity between Q and K, then scales and softmax normalizes the similarity, and finally uses the normalized weights to perform weighted summation on the V vector to obtain the output of attention.

[0031] Residual connections and layer normalization layers use residual connections to improve the gradient vanishing problem during the back propagation of deep networks, and layer normalization is used to transform the data into a standard distribution with a mean of 0 and a variance of 1.

[0032] The feed-forward fully connected layers perform nonlinear transformation on the multi-head attention output and use the ReLU function as the activation function to further extract features and increase nonlinearity.

[0033] Beneficial effects:

[0034] 1. Automatic feature extraction: The SSM-Transformer-based method of the present invention effectively reduces the reliance on manual feature extraction, and automatically characterizes the timing dependency information of the HRRP sequence through the SSM module, thereby improving the efficiency and accuracy of feature extraction.

[0035] 2. Extraction of deep temporal correlation features: Utilizing the multi-head attention mechanism in the Transformer encoder, the present invention can deeply mine the deep temporal correlation features in the sequence, thereby significantly improving the performance of passive interference recognition.

[0036] 3. Improve recognition accuracy: By comprehensively utilizing the timing dependency information and deep feature extraction through the SSM module and the Transformer encoder, the present invention achieves a higher recognition accuracy in the passive interference recognition task and provides more reliable technical support for applications in related fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 The overall framework diagram of the passive interference identification method based on SSM-Transformer;

[0038] Figure 2 Simulated HRRP for chaff clouds, corner reflector arrays, and ships;

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

[0040] Figure 4 This is a comparison chart of recognition accuracy performance of different network structures;

[0041] Figure 5 Recognition accuracy plots for datasets with different nuisance parameter distributions. DETAILED DESCRIPTION

[0042] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is clearly and completely described below in conjunction with the embodiments of the present application and the accompanying drawings.

[0043] The detailed process of this implementation method is as follows:

[0044] Figure 1 This is the overall framework diagram of the passive interference identification method based on SSM-Transformer. The overall process is divided into three parts: data set construction, preprocessing, and model construction. The preprocessing is divided into two-dimensional reconstruction and position encoding, and the model construction is divided into SSM module and Transformer encoder:

[0045] Step S1, constructing a data set by acquiring HRRP sequence based on chaff cloud, corner reflector array, and ship echo:

[0046] Through simulation, datasets of chaff cloud interference, corner reflector array interference and ship targets were constructed. For chaff interference, a dynamic model was first established for a single chaff, and the motion trajectory was simulated according to parameters such as wind speed, initial velocity and attitude angle distribution. The spatial distribution of multiple chaff strips was generated, and the chaff cloud interference echo was obtained by superposition. HRRP was obtained by pulse compression. For corner reflector array interference, linear, square and circular array configurations were constructed using octahedral corner reflectors and icosahedral corner reflectors with a side length of 1m. Different spacing parameters were set, and the RCS was calculated by physical optics algorithm, and IFFT conversion was performed to obtain the equivalent HRRP. The ship size is 80m×10m×15m, and the method is the same as that of the corner reflector array. The elevation 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 number of HRRP sampling points is 200. The normalized HRRP samples under different model parameters and angles obtained by simulation are as follows: Figure 2 shown.

[0047] Step S2, reconstruct the HRRP sequence in two dimensions, segment the HRRP into multiple row vectors, and form a sequence vector matrix X0, so that the distance between different sampling points that are far apart is reduced. The sequence vector matrix X0 can be expressed as:

[0048]

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

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

[0051]

[0052] Where t=1,2,…,D1 represents the position order, 2i and 2i+1 represent the even and odd dimensions from 1 to D2, respectively. The input matrix X is obtained by adding the sequence vector matrix X0 and its position encoding.

[0053] Step S4, construct an SSM module to extract the front-end dependency information of each subsequence:

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

[0055]

[0056] Among them, x k is the input vector, y k is the output vector, A, B, C are all system matrices, X is the input matrix, K is the convolution kernel, and y is the output matrix. Matrix A is related to the state vector, contains the information of the previous state to construct the new state, and it also determines how this information is transmitted in the time series. In order to better capture the information of subsequences with closer distances and attenuate the information with farther distances, high-order polynomial projection operators (HiPPO) are used to create matrix A. The HiPPO square matrix can be expressed as:

[0057]

[0058] Among them, n is the matrix row number, and k is the matrix column number. The convolution kernel K is used to convolve with the input to obtain the temporal relationship information between each row vector of the data, which is sent to the Transformer encoder to help extract richer temporal relationship features.

[0059] Step S5, use the multi-head attention layer in the Transformer encoder to extract features to obtain time series features for training and testing to complete passive interference identification. The HRRP samples are divided into training sets and test sets in a ratio of 8:2. During training, the input sample batch is set to 8, the training round is 200 rounds, and the learning rate is 0.001.

[0060] The multi-head attention layer is a structure that fuses the information extracted from multiple single-head self-attention layers. The structure diagram is as follows: Figure 3 As shown in the figure, the input data matrix is ​​transformed into Q , W K and W V Multiply them for linear transformation to get the three input matrices of the attention mechanism: query matrix Q (query), key matrix K (key) and value matrix V (value). Based on Q, the correlation between Q and all key matrices K is calculated by the inner product between the two to get a weight distribution. The greater the correlation, the greater the weight of the corresponding value matrix V, which indicates the importance of different element information. For the input sequence x, the output of the attention mechanism can be expressed as:

[0061]

[0062] in,(·) T represents matrix transpose, QK T Represents matrix Q and matrix K T Perform the dot product, d kRepresents the dimension size of matrix K. The formula first calculates the similarity between Q and K, then scales and normalizes the similarity using softmax, and finally uses the normalized weights to weight the sum of the V vectors to obtain the output of attention. Finally, the recognition result is output through the fully connected layer and the softmax function.

[0063] Without loss of generality, the accuracy metric is used to evaluate the recognition performance.

[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 recognition accuracy changes with JNR as shown in the figure below. Figure 4 As shown in (a), when a 6-layer encoder and different numbers of multi-head attention heads are used, the recognition accuracy changes with JNR. Figure 4 (b) as shown.

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

[0066] In order to verify that the proposed method is superior to other recognition methods based on deep learning, the Transformer, ResNet18 and LSTM without SMM modules are compared. Among them, the Transformer network structure uses a 10-layer encoder and 16 attention heads. At the same time, in order to verify that the proposed method has stronger generalization ability than other recognition methods based on deep learning, four data sets with different interference parameter distributions are simulated and constructed to analyze the changes in interference recognition performance. The types and values ​​of interference parameters in different data sets are shown in Table 1.

[0067] Table 1

[0068]

[0069] The neural network was trained using a training set with the same distribution as dataset a, and then tested using four datasets. For different recognition methods, the recognition accuracy and the decrease in recognition rate compared to dataset a were statistically analyzed. The results are as follows: Figure 5 shown.

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

[0071] In summary, the passive interference identification method based on SSM-Transformer proposed in this paper can fully extract the temporal relationship and improve the recognition performance.

[0072] The above description shows and describes several preferred embodiments of the present invention, but as mentioned above, it should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the invention concept described herein through the above teachings or the technology or knowledge of the relevant field. Changes and variations made by those skilled in the art do not depart from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.

Claims

1. A passive interference identification method based on HRRP and SSM-Transformer, characterized in that: The method comprises the following steps: Step S1, constructing a data set by acquiring HRRP sequence based on chaff cloud, corner reflector array, and ship echo; Step S2, reconstructing the HRRP sequence in two dimensions, segmenting the HRRP into multiple row vectors, and forming a sequence vector matrix X0, so that different sampling points that are far apart are shortened; Step S3, characterizing the order information between the row and column vectors in X0 based on the position coding, so that each row and column vector has a front-and-back relationship; Step S4, constructing an SSM module to extract the front-end dependency information of each subsequence; Step S5, using the Transformer encoder to extract features and obtain hidden layer expressions for classification.

2. The passive interference identification method based on HRRP and SSM-Transformer according to claim 1, characterized in that: The specific method of step S1 is: Through simulation, datasets of chaff cloud interference, corner reflector array interference and ship targets are constructed. For chaff interference, a dynamic model is firstly established for a single chaff and the motion trajectory is simulated according to parameters such as wind speed, initial velocity and attitude angle distribution to generate the spatial distribution of multiple chaffs. The echo of chaff cloud interference is obtained by superposition, and HRRP is obtained by pulse compression. For corner reflector array interference, octahedral corner reflectors and icosahedral corner reflectors with a side length of 1m are used to construct linear, square and circular array configurations, and different spacing parameters are set. The RCS is calculated by the physical optics algorithm, and the equivalent HRRP is obtained by IFFT conversion. The ship size is 80m×10m×15m, and the method is the same as that of the corner reflector array. The elevation 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 number of HRRP sampling points is 200.

3. The passive interference identification method based on HRRP and SSM-Transformer according to claim 1, characterized in that: The specific method of step S2 is: The HRRP sequence is reconstructed in two dimensions, and the HRRP is segmented into multiple row vectors to form a sequence vector matrix X0, so that the distance between different sampling points that are far apart is reduced; the sequence vector matrix X0 can be expressed as: Among them, x 1×N is HRRP, D1 is the total number of segments, and the length of each segment is D2=N / D1.

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

5. The passive interference identification method based on HRRP and SSM-Transformer according to claim 1, characterized in that: The calculation formula of the position code PE in step S3 is: Where t=1,2,…,D1 represents the position order, 2i and 2i+1 represent the even and odd dimensions from 1 to D2 respectively; the input matrix X is obtained by adding the sequence vector matrix X0 and its position code.

6. The passive interference identification method based on HRRP and SSM-Transformer according to claim 1, characterized in that: The expression of SSM in step S4 is as follows: and k =CA k Bx0+CA k-1 Bx1+...+CABx k-1 +CBx k y=X*K Among them, x k is the input vector, y k is the output vector, A, B, C are all 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 the new state, and it also determines how this information is transmitted in the time series; in order to better capture the information of subsequences with a closer distance and attenuate the information with a farther distance, the high-order polynomial projection operator (HiPPO) is used to create the matrix A. The HiPPO square matrix can be expressed as: Among them, n is the matrix row number, and k is the matrix column number. The convolution kernel K is used to convolve with the input to obtain the temporal relationship information between each row vector of the data, which is sent to the Transformer encoder to help extract richer temporal relationship features.

7. The passive interference identification method based on HRRP and SSM-Transformer according to claim 1, characterized in that: The Transformer encoder described in step S5 mainly includes a multi-head attention layer, a residual connection and layer normalization, and a feedforward fully connected layer.

8. The passive interference identification method based on HRRP and SSM-Transformer according to claim 1, characterized in that: In step S5, the multi-head attention layer in the Transformer encoder is used to extract features to obtain time series features for training and testing to complete passive interference identification; the HRRP samples are divided into training sets and test sets in a ratio of 8:2; during training, the input sample batch is set to 8, the training rounds are 200 rounds, and the learning rate is 0.

001.

9. The passive interference identification method based on HRRP and SSM-Transformer according to claim 1, characterized in that: In step S5, for the input sequence x, the output of the attention mechanism is expressed as: Q=x·W Q ,K=x·W K ,V=x·W V in,(·) T represents matrix transpose, QK T Represents matrix Q and matrix K T Perform the dot product, d k Represents the dimension size of the matrix K; the formula first calculates the similarity between Q and K, then scales and normalizes the similarity, and finally uses the normalized weights to weighted sum the V vector to obtain the output of attention; finally, the recognition result is output through the fully connected layer and the softmax function.

Images