Radar target identification method based on polarized high-resolution range profile sequence

Abstract

The invention provides a feature fusion method based on a high resolution range profile (HRRP) sequence, which is used for radar automatic target identification, and aims at fusing various features in the high resolution range profile sequence so as to improve the identification performance. The invention designs a novel network framework for extracting and integrating various features from a multi-polarization HRRP sequence. The network is divided into four key components: a Tokenization module, a feature extraction module, a feature fusion module and a classification module.

Description

Technical Field

[0001] The invention relates to the field of radar automatic target recognition, and in particular to a target classification method based on a polarization high-resolution range image sequence. Background Art

[0002] Radar Automatic Target Recognition (RATR) technology was originally developed for military applications and is now widely used in civilian and military fields. Existing RATR methods can be divided into three categories according to the type of radar data: Radar Cross Section (RCS)-based, HRRP-based, and Synthetic Aperture Radar (SAR) image or Inverse Synthetic Aperture Radar (ISAR) image-based methods. Although RCS contains scattering information, it is not easy to extract the structural features of the target from it. Recognition based on SAR and ISAR imaging is limited by the complexity of the imaging algorithm, its dependence on target motion, and its sensitivity to electromagnetic interference and noise. HRRP is formed by the coherence of radar echoes along the radar line of sight (RLOS), and contains the structural information of the target scattering center and the size information in the RLOS direction. Compared with RCS data, HRRP provides more intuitive target structure information. Compared with ISAR images, HRRP is obtained by pulse compression of broadband radar echoes in the range direction, which is easier to acquire, store and process, while still providing richer information. In addition, HRRP sequences contain time information, which can be used to track the changes of target structures over time, providing a dynamic perspective for accurate and reliable RATR in different operating environments.

[0003] Existing RATR methods based on HRRP can be divided into two categories. The first category is based on traditional signal processing, such as using features such as bispectrum and high-order spectrum or using statistical models to extract statistical features of the target for classification and identification. However, this type of method is overly dependent on prior knowledge of the target. The second category is machine learning methods. With the development of artificial intelligence technology, more and more deep learning networks such as LSTM and Transformer have been applied to this field and have achieved good results.

[0004] Polarization information is another important feature for characterizing targets. Commonly used polarizations are HH, HV, VV, and VH polarizations. Electromagnetic waves with different polarization modes respond differently to the target being measured, thus providing different information for target identification. This polarization diversity significantly enhances the radar's ability to collect more information about the target structure.

[0005] Each polarization channel contains global features and local features. Current methods often only focus on the local features of HRRP or the temporal relationship and polarization information of the HRRP sequence alone, which leads to low recognition accuracy and poor model robustness. The existing feature fusion methods are single and cannot effectively fuse the polarization global features, local features and HRRP sequences. Fusion of the two can improve recognition accuracy and robustness. Therefore, it is crucial to fully utilize the temporal and polarization information of HRRP to more comprehensively represent the target characteristics and develop an algorithm that can effectively extract and fuse global and local features from multiple polarization modes. Summary of the invention

[0006] The present invention aims to provide a radar automatic target recognition algorithm with good robustness by fusing the global and local features of HRRP under different polarizations. To achieve the above purpose, this paper proposes a two-stage feature fusion method, which fuses the global and local features of different polarizations respectively, and significantly improves the results of the classification task. Based on the above content, a multi-polarization feature fusion network is implemented to extract multiple features from the polarization HRRP sequence. The network is mainly composed of a Tokenization module, a feature extraction module, a feature fusion module and a classification module. The method comprises the following steps:

[0007] S1: Collect polarized HRRP data, process the raw data to obtain a data set, and divide the data set into a test set and a training set; the collected data set includes multiple polarized HRRP data; divide the collected data into several HRRP data sequences according to the time series relationship, and annotate each data sample with a category label;

[0008] S2: Modulo 2 norm normalization is used to normalize the amplitude of the real high-resolution range image data to obtain a normalized high-resolution range image sequence. This method can solve the amplitude sensitivity problem in HRRP data.

[0009] S3: The Tokenization module first encodes the HRRP sequence through a two-layer convolutional neural network and realizes the preliminary feature extraction. Then, the encoded data is activated by ReLU and Max Pooling to reduce the data dimension but retain the significant features of the sequence. Then, the position encoder layer is used to embed the position information into the code corresponding to the HRRP sequence. Finally, the encoded result is output to the feature extraction module for feature extraction.

[0010] S4: The feature extraction module is used to further extract feature information in the HRRP sequence. The feature extraction module consists of two submodules: global feature extraction and local feature extraction. The global feature extraction submodule is designed based on the Transformer encoder and is used to extract the spatiotemporal correlation in the HRRP sequence. The local feature extraction module is designed based on the residual CNN layer and is used to extract the local scattering center features in HRRP. Finally, the feature extraction module outputs the extracted global features and local features to the feature fusion module respectively.

[0011] S5: The features of the HRRP sequences under different polarizations are fused using a feature fusion module. The feature fusion module is a two-stage fusion module. In the first stage of the feature fusion module, the global features and local features under different polarizations are fused separately through a gated fusion module; in the second stage, the global features and local features of different polarizations are fused through a cross-attention module. So far, the present invention has completed all feature extraction and fusion work.

[0012] S6: Use the classification module to classify the fused features. The classification module consists of a fully connected layer and a softmax layer. The fully connected layer reduces the dimension of the fused features, and the softmax layer produces the final classification prediction. This module outputs the classification results to complete the classification and recognition of the target.

[0013] Compared with the prior art, the technical solution of the present invention has the following significant beneficial effects:

[0014] (1) In the embodiment of the present invention, the global module and the local module are organically combined to significantly enhance the network's ability to extract HRRP features and effectively improve the network's recognition accuracy.

[0015] (2) The polarization information of the HRRP sequence is used in the embodiment of the present invention. The features of the HRRP sequence under different polarizations are organically combined to form a more comprehensive feature expression, effectively improve the recognition accuracy of the network, and enhance the robustness of the network.

[0016] (3) The feature extraction module in the embodiment of the present invention adopts the Transformer encoder structure, because the Transformer encoder structure can capture a wide range of global features and provide feature information of a global perspective for target recognition.

[0017] (4) The feature extraction module in the embodiment of the present invention uses a residual CNN layer. This is because the residual CNN layer can effectively extract fine-grained local features of the HRRP sequence, allowing the model to focus on the local features of the target, which will complement the global information in the Transformer encoder and form a more comprehensive information expression for the HRRP sequence.

[0018] (5) The feature fusion module in the embodiment of the present invention adopts a gating mechanism, which organically integrates the features of the HRRP sequence by adaptively adjusting the weights of information under different polarizations, thereby enhancing the generalization performance of the network.

[0019] (6) The feature fusion module in the embodiment of the present invention adopts a cross-attention mechanism. The cross-attention mechanism forms a more comprehensive feature expression by adaptively calculating the relationship between global and local information, effectively improving the recognition accuracy of the network and enhancing the robustness of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component shown in various figures may be represented by the same reference numeral. For clarity, not every component is labeled in every figure. Embodiments of various aspects of the present invention will now be described by way of example and with reference to the accompanying drawings, in which:

[0021] Figure 1 is a summary drawing of the present invention;

[0022] Figure 2 is a flow chart of the target recognition and classification method of the present invention;

[0023] Figure 3 It is the Tokeniztion module;

[0024] Figure 4 It is a global feature extraction module;

[0025] Figure 5 It is a local feature extraction module;

[0026] Figure 6 It is a multi-head attention mechanism;

[0027] Figure 7 It is the gated fusion module;

[0028] Figure 8 It is a cross-attention mechanism;

[0029] Fig. 9 It is the target recognition accuracy of different networks under different signal-to-noise ratio conditions. DETAILED DESCRIPTION

[0030] In order to better understand the technical content of the present invention, a specific embodiment is given and illustrated as follows. The method proposed in the present invention can theoretically be applied to any multi-polarization HRRP target recognition, such as satellite recognition, human recognition, multi-target fine-grained recognition, etc. In different application scenarios, it is only necessary to change the category label of the classification target and retrain. Figure 2The flowchart is a step flow chart of the fine-grained classification of satellite targets in the radar target recognition method based on polarization high-resolution range profile sequence according to an embodiment of the present invention, which includes the following steps:

[0031] S1: This embodiment uses a data set obtained by simulation to extract radar echo signals of the target in the HH and VH polarization channels at multiple consecutive moments;

[0032] S2: Extract HRRP from the radar echo signals under two polarizations respectively, and arrange them in the time dimension to form a polarization HRRP sequence. Divide the HRRP sequence into a test set and a training set;

[0033] S3: Process the HRRP data using modulo 2 norm normalization to obtain the normalized HRRP sequence. Input the processed HH and VH polarized HRRP sequences into the Tokenization module respectively;

[0034] S4: The Tokenization module processes the input HRRP sequence. This module first performs preliminary encoding and feature extraction on the HRRP sequence through CNN, then further processes the encoded data through the position encoder layer, and finally outputs the encoded result to the feature extraction module for feature extraction;

[0035] S5: The feature extraction module extracts the global and local features of the target through global feature extraction and local feature extraction, and outputs the extracted global features and local features to the feature fusion module;

[0036] S6: The feature fusion module first fuses the global features and local features of different polarizations, then fuses the global features and local features, and finally outputs the fusion results to the classification module;

[0037] S7: The classification module completes the classification and recognition of the target based on the fused features;

[0038] S8: Evaluate the results.

[0039] In a specific application embodiment, the detailed steps of S2 are:

[0040] S201: During the model training process, the optimal model is obtained by fixing the parameters. The training data set is input into the model, and the model starts forward propagation. After completion, the output category is compared with the input category, and the loss function is calculated. Then the model starts backward propagation, and the weights of each layer of the model are adjusted using the gradient descent method. After that, the forward propagation is repeated until the loss function is small.

[0041] In a specific application embodiment, the detailed steps of S3 are:

[0042] S301: Assume that the acquired HRRP data is s=[d1, d2, ..., d L ] T , where d i represents the sub-echo of the ith range unit, and L represents the total number of range units. By performing a modulo operation on the complex HRRP data, the amplitude of the HRRP data can be obtained as s = |s| = [d1, d2, ..., d L ] T The HRRP amplitude data s is normalized and expressed as:

[0043]

[0044] In a specific application embodiment, the detailed steps of S4 are:

[0045] S401: As reference Figure 3 As shown in the figure, the Tokenization module first performs preliminary encoding and feature extraction on the HRRP sequence through two layers of CNN. The results after convolution and pooling are activated by ReLU (Rectified Linear Unit). The ReLU activation function can introduce nonlinear characteristics, allowing the model to capture complex nonlinear relationships and avoid the gradient vanishing problem, thereby improving the learning ability of the model. Suppose the input HRRP sequence is X, then:

[0046] Z0=ReLU(Pool(Conv(X,W T0 ,b T0 )))

[0047] Z1=ReLU(Pool(Conv(Z0,W T1 ,b T1 )))

[0048] Where W T0 , W T1 are the convolution kernels of the first and second layers, b T0 , b T1 are the bias terms of the first and second layers respectively.

[0049] S402: The Tokenization module then further processes the encoded data through the position encoder layer. Due to the sequential nature of the HRRP sequence itself, the task of the position encoder layer is to embed the position information into the encoding result of each time step. The specific process is shown in the following formula:

[0050] Z2=Z1+P

[0051] Among them, P is the position encoding matrix, and Z2 is the result after position encoding processing, which contains the sequential information of time steps, enabling the model to learn the temporal dynamics of target motion.

[0052] In a specific application embodiment, the detailed steps of S5 are:

[0053] S501: As reference Figure 4 As shown in the figure, global feature (temporal relationship between sequences) extraction is implemented through a Transformer-based encoder. The output of the Tokenization module first passes through a network with a multi-head self-attention mechanism, and the network output is processed by a position-dependent FFN, which has two linear transformations and a GeLU activation function in the middle. This structure enhances the model's ability to learn complex representations by introducing nonlinearity and allows for a deeper understanding of the relationships within the feature space. It is shown in the following formula:

[0054]

[0055]

[0056] In the formula is the final output of global feature extraction for any polarization type, is the learnable weight matrix, is the corresponding bias, is a scaling factor used to ensure the stability of the softmax function, W G , b G are weights and biases.

[0057] S502: As reference Figure 5 As shown in the figure, local feature (fine-grained feature) extraction is achieved through residual CNN layer. CNN block can be used to capture detailed information of the target scattering center, and residual connection enables the model to extract local features, which are then added to the input features to facilitate the flow of gradient information during training. The specific process is as follows:

[0058]

[0059] in, is the final output of local feature extraction of any polarization type, They are the first and second convolution kernels respectively. are the first and second layer bias terms respectively.

[0060] In a specific application embodiment, the detailed steps of S6 are:

[0061] S601: The first stage of the feature fusion module fuses the global features and local features of different polarizations respectively. Figure 7 As shown in Figure 2, global feature fusion is implemented through a gated fusion module. The specific process is shown in the following formula:

[0062]

[0063] in and are the global features of HH polarization and VH polarization, W1, W2, W3 are learnable weight matrices, b1, b2, b3 are biases associated with the linear layer, are the hidden layer outputs of HH polarization and VH polarization respectively. The output of the global feature fusion block can be expressed as:

[0064]

[0065] where α g is the fusion weight of two types of polarization features.

[0066] S602: Local feature fusion is implemented through a gated fusion module, and the specific process is shown in the following formula:

[0067]

[0068] in and are the local fine-grained features of HH polarization and VH polarization, W4, W5, W6 are learnable weight matrices, b4, b5, b6 are biases associated with the linear layer, are the hidden layer outputs of HH polarization and VH polarization respectively. The output of the global feature fusion block can be expressed as:

[0069]

[0070] where α d is the fusion weight of two types of polarization features.

[0071] S603: The second stage of the feature fusion module is the fusion of global features and local features, which is achieved through cross attention blocks, as shown in Figure 8 As shown. The input of the cross attention block is the output F of the global feature fusion block and the local feature fusion block. G 、F D , using the cross self-attention mechanism, the specific process is shown in the following formula:

[0072] Q=W Q ·F G +b Q

[0073] K=W K·F D +b K

[0074] V=W V ·F D +b V

[0075]

[0076] Where F is the final output of feature fusion, W Q , Q K , W V is the learnable weight matrix, b Q , b K , b V is the corresponding bias, is a scaling factor used to ensure the stability of the softmax function, and W and b are weights and biases.

[0077] In a specific application embodiment, the detailed steps of S7 are:

[0078] S701: The classification module is implemented through a network based on a fully connected layer and a softmax layer. The fully connected layer integrates the overall features and maps the features to the classification space. The softmax layer generates the final classification score and outputs the final prediction result. This module outputs the classification result to complete the classification and recognition of the target.

[0079] In a specific application embodiment, the detailed steps of S8 are:

[0080] S801: When evaluating the algorithm, the overall recognition performance of the model is analyzed by simulation experiments, and then the recognition accuracy of the proposed algorithm is compared with that of other algorithms using test data under different signal-to-noise ratio conditions to verify the feasibility and advantages of the algorithm.

[0081] S802: By Fig. 9It can be seen that the algorithm proposed by the present invention shows extremely high accuracy under high signal-to-noise ratio conditions. When the signal-to-noise ratio is 20dB, its accuracy reaches about 93%, which is significantly better than other algorithms. Among them, the accuracy of the Transformer method under the same signal-to-noise ratio is about 88%, the LSTM method is about 85%, and the traditional CNN method only reaches about 78%. This shows that the proposed method can fully tap the data features and improve the classification performance. In addition, under low signal-to-noise ratio conditions (such as when the signal-to-noise ratio is 0dB and 5dB), the proposed method also has significant robustness, and the accuracy is always ahead of other algorithms. When the signal-to-noise ratio is 0dB, the accuracy of the proposed method is about 73%, which is significantly improved compared to 67% of the LSTM method, 65% of the Transformer method, and 59% of the CNN method. This further proves the ability of the proposed method to suppress noise interference in complex environments. The proposed algorithm shows excellent classification performance and robustness under the full range of signal-to-noise ratio conditions, especially in terms of classification accuracy under high signal-to-noise ratio conditions and anti-interference ability under low signal-to-noise ratio conditions, which provides strong support for signal classification tasks in complex environments in practical applications.

[0082] Various aspects of the present invention are described in this disclosure with reference to the accompanying drawings, in which illustrative embodiments are shown. The embodiments of the present disclosure are not necessarily defined to include all aspects of the present invention. It should be understood that the concepts and embodiments introduced above, as well as those described in more detail below, can be implemented in any of many ways, because the concepts and embodiments disclosed by the present invention are not limited to any implementation. In addition, some aspects disclosed by the present invention can be used alone, or in any appropriate combination with other aspects disclosed by the present invention.

[0083] Although the present invention has been disclosed as above with preferred embodiments, it is not intended to limit the present invention. A person with ordinary knowledge in the technical field to which the present invention belongs may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be determined by the definition of the claims.

Claims

1. A radar target recognition method based on polarization high-resolution range profile sequence, characterized by: Adaptively perform polarization feature fusion. The method comprises: S1: Get the multi-polarized HRRP sequence of the target and use it as the input of the network; S2: The Tokenization module converts the continuous polarized HRRP sequence into word segmentation representation, preparing for structured processing in the subsequent stage; S3: The feature extraction module extracts polarization feature information from the multi-polarization HRRPs respectively; S4: The feature fusion module will adaptively fuse the multi-polarization feature information to obtain the comprehensive feature information of the target; S5: The classification module classifies the comprehensive feature information of the target to obtain the category information of the target.

2. The radar target recognition method based on polarization high-resolution range profile sequence according to claim 1, characterized in that: In step S1, the input of the method is the multi-polarized HRRP time series data of the target, so as to extract more comprehensive target features.

3. The radar target recognition method based on polarization high-resolution range profile sequence according to claim 1, characterized in that: The Tokenization module in step S2 consists of two convolutional neural networks (CNNs) and a position encoding layer. The module finally outputs a preliminarily processed sequence to the feature extraction module.

4. The radar target recognition method based on polarization high-resolution range profile sequence according to claim 2, characterized in that: Two CNN layers are used to capture the spatial information inherent in the HRRP sequences, enabling the model to learn the spatial patterns associated with the target scattering centers.

5. The radar target recognition method based on polarization high-resolution range profile sequence according to claim 2, characterized in that: The input of the position encoding layer is the output of the CNN layer, which encodes the position information and captures the temporal relationship between adjacent sequences, allowing the model to learn the dynamic changes of the object motion.

6. The radar target recognition method based on polarization high-resolution range profile sequence according to claim 1, characterized in that: The feature extraction module in step S3 includes a global feature extraction branch and a local feature extraction branch.

7. The radar target recognition method based on polarization high-resolution range profile sequence according to claim 6, characterized in that: The input of the global feature extraction branch is the sequence generated by the Tokenization module, which is a Transformer encoder.

8. The radar target recognition method based on polarization high-resolution range profile sequence according to claim 6, characterized in that: The input of the local feature extraction branch is the sequence generated by the Tokenization module. This branch consists of several stacked residual CNN layers to extract fine-grained local features.

9. The radar target recognition method based on polarization high-resolution range profile sequence according to claim 1, characterized in that: Step S4 The feature fusion module is a two-stage data processing process.

10. The radar target recognition method based on polarization high-resolution range profile sequence according to claim 9, characterized in that: The first stage of the feature fusion module in step S4 includes a global feature fusion block and a local feature fusion block. The input of the fusion module is the global features and local features extracted in step S3. The global feature fusion block and the local feature fusion block have the same network structure, which is composed of gated multimodal units (GMU).

11. The radar target recognition method based on polarization high-resolution range profile sequence according to claim 9, characterized in that: The second stage of the feature fusion module in step S4 is the cross attention block. The input of this module is the output of the global feature fusion block and the local feature fusion block. It is used to fuse global features and local features to form a comprehensive expression of the target features.

12. The radar target recognition method based on polarization high-resolution range profile sequence according to claim 1, characterized in that: The classification module in step S5 includes a fully connected layer and a softmax activation function, which is used to give the target classification result.

Images