HRRP radar target identification method based on Inception-GAN
The timing and spatial structure characteristics of HRRP data are extracted and integrated through the Inception-GAN model, combined with the deep reinforcement learning framework, the problem of insufficient timing feature extraction in the existing technology is solved, and a more efficient HRRP target recognition effect is achieved.
Patent Information
- Application Number
- CN202411726145.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-28
- Publication Date
- 2025-05-02
AI Technical Summary
The existing HRRP target recognition methods mainly rely on timing feature extraction, and fail to fully utilize the spatial structure features in HRRP data, resulting in unsatisfactory recognition results.
Using the Inception-GAN-based method, the timing and spatial structure characteristics of HRRP data are extracted by improving the one-dimensional convolution structure and graph attention convolution network, and the two are fused to build a deep reinforcement learning framework to improve the recognition performance of the model.
It achieves a better HRRP target recognition effect under a lower parameter quantity, improves the stability and recognition accuracy of the model, and accelerates the convergence speed of the model.
Smart Images
Figure CN119919626A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of radar target recognition, relates to HRRP radar target recognition technology, and specifically relates to an HRRP radar target recognition method based on Inception-GAN. Background Art
[0002] In the radar automatic target recognition system, HRRP contains a wealth of target structure features, such as target size and scatterer distribution. At the same time, HRRP uses vector form, which effectively reduces the amount of storage and calculation. Therefore, HRRP is a research hotspot in the field of radar automatic target recognition.
[0003] Traditional HRRP target recognition is to extract features with specific physical meanings from HRRP, such as power spectrum, cepstrum, FFT amplitude spectrum and other features. This kind of artificially extracted features is effective, but its quality depends on personal experience to a certain extent, which limits their use in practice. Due to the nonlinearity of deep learning, it can automatically learn different levels of features from HRRP data, so as to have better recognition performance, so HRRP target recognition methods based on deep learning are gradually emerging. Since HRRP is a one-dimensional projection of three-dimensional target scattering along the radial direction, there is a time correlation between the distance units within the HRRP sample. Therefore, many studies mainly achieve target classification by extracting time series features from HRRP data. However, HRRP data contains structural feature information, and the potential hierarchical structure can provide more useful information for sequence classification. Therefore, only extracting the time series features of HRRP, the recognition effect is not very ideal. Summary of the invention
[0004] Purpose of the invention: In order to overcome the shortcomings of the prior art, a HRRP radar target recognition method based on Inception-GAN is provided, which can simultaneously extract the temporal features and spatial structure features of HRRP data, and fuse the extracted features. At the same time, by utilizing the constructed deep reinforcement learning framework, the constructed model can achieve better HRRP target recognition effect with a lower number of parameters, and at the same time can improve the convergence speed of the model.
[0005] Technical solution: To achieve the above purpose, the present invention provides a HRRP radar target recognition method based on Inception-GAN, comprising the following steps:
[0006] S1: preprocess the original data set and convert the data set into HRRP sequence;
[0007] S2: Use the constructed Inception-GAN network model to extract the temporal features and spatial features of the preprocessed HRRP sequence, and fuse the two features;
[0008] S3: Construct the Inception-GAN deep reinforcement learning framework based on the feature extraction capability of the Inception-GAN model and the fused features;
[0009] S4: Train the Inception-GAN network model based on the Inception-GAN deep reinforcement learning framework to output the optimal recognition results.
[0010] Furthermore, the specific process of step S1 is as follows:
[0011] The original data set is subjected to one-dimensional inverse fast Fourier transform to obtain the complex domain HRRP sequence, and then its amplitude is taken and normalized.
[0012] Furthermore, in step S1, the amplitude of the complex domain HRRP sequence is taken, and its expression is:
[0013]
[0014] where S = {s1,s2,...,s N} is the HRRP sequence of the complex domain; s n =a n +jb n For each complex sampling point;
[0015] The linear normalization method is used to normalize the amplitude of the HRRP after taking the amplitude, and its expression is:
[0016]
[0017] where |s' n | is the normalized amplitude; min(|S|) and max(|S|) are the minimum and maximum amplitudes of the HRRP sequence, respectively.
[0018] Furthermore, the construction of the Inception-GAN network model in step S2 includes: improving the one-dimensional convolution structure, wherein the improved one-dimensional convolution structure extracts different fine-grained features of the HRRP sequence by using three convolution kernels of different lengths, and down-samples the features by a parallel pooling convolution layer, so as to reduce the feature dimension after the improved one-dimensional convolution structure extracts the temporal features;
[0019] Add a channel attention mechanism after the improved one-dimensional convolution structure;
[0020] Design an improved graph attention convolutional network to extract spatial structure features.
[0021] Furthermore, the spatial feature extraction method in step S2 is:
[0022] In order to extract the spatial structural features of the HRRP feature sequences, the spatial structural information between the HRRP feature sequences needs to be represented by an adjacency matrix. First, the spatial distance between any two feature vectors is calculated by calculating the Euclidean distance matrix, and its expression is:
[0023]
[0024] Where D jk Represents X i The Euclidean distance between the feature vectors of the j-th and k-th nodes in , ||·||2 represents the Euclidean norm;
[0025] The adjacency matrix is constructed based on the Euclidean distance matrix, and its expression is:
[0026]
[0027] Among them A i,jk Represents X i The connectivity relationship between the feature vectors of the j-th and k-th nodes in ;
[0028] Then after obtaining the adjacency matrix, the adjacency matrix A i The node feature matrix X is used as the input of the improved graph attention convolutional network (GAN) to extract spatial structure features, which is expressed as:
[0029]
[0030] where Λ i ={j|A i,jk >0}, W is the linear transformation matrix;
[0031]
[0032] Among them, α ij (k) is the weight of each attention head k.
[0033] Furthermore, in step S2, the extracted temporal features and spatial features are fused by a hybrid attention mechanism to obtain a fused feature, which is expressed as:
[0034] F fusion =softmax(W f [F1||F2])
[0035] Among them, W f is the learnable weight matrix.
[0036] Furthermore, in step S3, in order to improve the convergence speed of the constructed model, the constructed Inception-GAN model is used as the feature extraction layer of the actor-critic network to construct the Inception-GAN deep reinforcement learning framework. During the construction process, the preprocessed HRRP sequence is used as the environment, and then the fused features obtained by the Inception-GAN model are input into the actor-critic network, so that the actor-critic network can interact with the fused features. At the same time, the classification features finally output by the construction framework can be continuously optimized by using the actor-critic network strategy update, thereby improving the accuracy of target recognition.
[0037] Furthermore, in step S3, after constructing the Inception-GAN deep reinforcement learning framework, in order to enable the feature vector obtained after the next batch of data passes through the feature extraction layer to guide the PPO agent to correctly select the next action, the convergence speed of the entire framework is improved by improving the hybrid loss function, which is expressed as:
[0038]
[0039] Among them, α, β, and γ are the weighting factors of the strategy function, value function, and loss function of the constructed Inception-GAN network, respectively.
[0040] Furthermore, in step S4, the fused temporal features and spatial structural features are input into the constructed Inception-GAN deep reinforcement learning framework, and an optimized actor-critic network is obtained through training with a PPO agent, wherein the actor network outputs an optimized strategy, i.e., the probability distribution of selecting an action in each state; the critic network outputs an optimized state value estimate for evaluating the potential value of a given state;
[0041] Under the framework of the agent and actor-critic network, the final classification result is output by selecting the best action in the optimal strategy, which is expressed as:
[0042] C = argmax(π(A t |S t ))
[0043] Among them, argmax is to select the category with the maximum probability in the probability distribution.
[0044] Beneficial effects: Compared with the prior art, the present invention utilizes the constructed Inception-GAN model to extract the temporal features and spatial structural features of the HRRP target and fuses the two to enhance the feature expression ability and thus improve the stability of the model training process; and uses the constructed model as the feature extraction layer of the actor-critic network and designs a hybrid loss function of the constructed deep reinforcement learning framework to accelerate the convergence speed of the entire framework; at the same time, the actor-critic network strategy update can continuously optimize the classification features of the final output of the construction framework, thereby improving the accuracy of target recognition. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 is a flow chart of the method of the present invention;
[0046] Figure 2 A structural diagram of the improved one-dimensional convolution in the present invention;
[0047] Figure 3 This is a structural diagram of the improved graph attention convolution in the present invention;
[0048] Figure 4 It is a schematic diagram of the structure of the deep reinforcement learning framework in the present invention;
[0049] Figure 5 The figure is the overall architecture diagram of the method of the present invention. DETAILED DESCRIPTION
[0050] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.
[0051] like Figure 1 and Figure 5 As shown, the present invention provides a HRRP radar target recognition method based on Inception-GAN, comprising the following steps:
[0052] S1: preprocess the original data set and convert the data set into HRRP sequence;
[0053] Perform one-dimensional inverse fast Fourier transform on the original data set to obtain the complex domain HRRP sequence, then take its amplitude and perform amplitude normalization;
[0054] Taking the amplitude of the complex domain HRRP sequence, its expression is:
[0055]
[0056] where S = {s1,s2,...,s N} is the HRRP sequence of the complex domain; s n =a n +jb n For each complex sampling point;
[0057] The linear normalization method is used to normalize the amplitude of the HRRP after taking the amplitude, and its expression is:
[0058]
[0059] where |s' n | is the normalized amplitude; min(|S|) and max(|S|) are the minimum and maximum amplitudes of the HRRP sequence, respectively.
[0060] S2: Use the constructed Inception-GAN network model to extract the temporal features and spatial features of the preprocessed HRRP sequence, and fuse the two features;
[0061] The construction of the Inception-GAN network model includes: improving the one-dimensional convolution structure. The improved one-dimensional convolution structure uses three convolution kernels of different lengths to extract different fine-grained features of the HRRP sequence, and downsamples the features in parallel through the pooling convolution layer to reduce the feature dimension after the improved one-dimensional convolution structure extracts the time series features, thereby reducing the network parameters. Figure 2 As shown, the improved one-dimensional convolution structure in this embodiment has three parallel one-dimensional convolutions, the number of characteristic channels is 32, and the convolution kernel lengths are 40, 20, and 10 respectively;
[0062] A channel attention mechanism is added after the improved one-dimensional convolution structure. In the channel attention mechanism, an improved graph attention convolutional network is designed to extract spatial structure features.
[0063] like Figure 3 As shown, the method for extracting spatial features is as follows: In order to extract the spatial structural features of the HRRP feature sequences, the spatial structural information between the HRRP feature sequences needs to be represented by an adjacency matrix; first, the spatial distance between any two feature vectors is calculated by calculating the Euclidean distance matrix, and its expression is:
[0064]
[0065] Where D jk Represents X i The Euclidean distance between the feature vectors of the j-th and k-th nodes in , ||·||2 represents the Euclidean norm;
[0066] The adjacency matrix is constructed based on the Euclidean distance matrix, and its expression is:
[0067]
[0068] Among them A i,jk Represents X i The connectivity relationship between the feature vectors of the j-th and k-th nodes in ;
[0069] Then after obtaining the adjacency matrix, the adjacency matrix A i The node feature matrix X is used as the input of the improved graph attention convolutional network (GAN) to extract spatial structure features, which is expressed as:
[0070]
[0071] where Λ i ={j|A i,jk >0}, W is the linear transformation matrix;
[0072]
[0073] where α ij (k) is the weight of each attention head k;
[0074] See also Figure 3 , the temporal features extracted by the improved one-dimensional convolution structure are input into the Transformer structure of the channel attention mechanism to extract the global features of any position in the sequence, and finally the extracted temporal features and spatial features are fused through the hybrid attention mechanism to obtain the fused features, which are expressed as:
[0075] F fusion =softmax(W f [F1||F2])
[0076] Where W f is a learnable weight matrix;
[0077] S3: Construct the Inception-GAN deep reinforcement learning framework based on the feature extraction capability of the Inception-GAN model and the fused features;
[0078] In order to improve the convergence speed of the constructed model, the constructed Inception-GAN model is used as the feature extraction layer of the actor-critic network to construct the Inception-GAN deep reinforcement learning framework, such as Figure 4As shown in the figure, during the construction process, the preprocessed HRRP sequence is used as the environment, and then the fusion features obtained by the Inception-GAN model are input into the actor-critic network, so that the actor network can output a more accurate action probability distribution in a given state based on the extracted high-dimensional features. The agent selects actions through the action probability distribution, and the critic network evaluates the state-action pair output reward and feeds it back to the agent. At the same time, the actor-critic network strategy update can continuously optimize the classification features finally output by the construction framework, thereby improving the accuracy of target recognition;
[0079] After building the Inception-GAN deep reinforcement learning framework, in order to make the feature vector obtained after the next batch of data passes through the feature extraction layer to guide the agent (PPO agent) to correctly select the next action according to the probability distribution output by the actor network, the convergence speed of the entire framework is improved by improving the hybrid loss function, which is expressed as:
[0080]
[0081] Among them, α, β, and γ are the weighting factors of the strategy function, value function, and loss function of the constructed Inception-GAN network, respectively.
[0082] S4: Train the Inception-GAN network model based on the Inception-GAN deep reinforcement learning framework to output the optimal recognition results;
[0083] like Figure 5 As shown in the figure, the fusion features of the fusion of temporal features and spatial structural features are input into the constructed Inception-GAN deep reinforcement learning framework, and an optimized actor-critic network is obtained through the training of the intelligent agent, so that the intelligent agent can maximize the cumulative return in the target classification task, wherein the actor network outputs an optimized strategy, that is, the probability distribution of selecting actions in each state; the critic network outputs an optimized state value estimate, which is used to evaluate the potential value of a given state;
[0084] Under the framework of the agent and actor-critic network, the final classification result is output by selecting the best action in the optimal strategy, which is expressed as:
[0085] C = argmax(π(A t |S t ))
[0086] Here argmax is the category with the maximum probability in the selected probability distribution.
Claims
1. A HRRP radar target recognition method based on Inception-GAN, characterized in that: The steps include: S1: preprocess the original data set and convert the data set into HRRP sequence; S2: Use the constructed Inception-GAN network model to extract the temporal features and spatial features of the preprocessed HRRP sequence, and fuse the two features; S3: Construct the Inception-GAN deep reinforcement learning framework based on the feature extraction capability of the Inception-GAN model and the fused features; S4: Train the Inception-GAN network model based on the Inception-GAN deep reinforcement learning framework to output the optimal recognition results.
2. According to claim 1, a HRRP radar target recognition method based on Inception-GAN is characterized in that: The specific process of step S1 is as follows: The original data set is subjected to one-dimensional inverse fast Fourier transform to obtain the complex domain HRRP sequence, and then its amplitude is taken and normalized.
3. According to claim 2, a HRRP radar target recognition method based on Inception-GAN is characterized in that: In step S1, the amplitude of the complex domain HRRP sequence is taken, and its expression is: where S = {s1,s2,...,s N } is the HRRP sequence of the complex domain; s n =a n +jb n For each complex sampling point; The linear normalization method is used to normalize the amplitude of the HRRP after taking the amplitude, and its expression is: where |s' n | is the normalized amplitude; min(|S|) and max(|S|) are the minimum and maximum amplitudes of the HRRP sequence, respectively.
4. The HRRP radar target recognition method based on Inception-GAN according to claim 1 is characterized in that: The construction of the Inception-GAN network model in step S2 includes: improving the one-dimensional convolution structure, wherein the improved one-dimensional convolution structure extracts features of different fine-grainedness of the HRRP sequence by using three convolution kernels of different lengths, and down-samples the features by a parallel pooling convolution layer, so as to reduce the feature dimension after the improved one-dimensional convolution structure extracts the time series features; Add a channel attention mechanism after the improved one-dimensional convolution structure; Design an improved graph attention convolutional network to extract spatial structure features.
5. According to claim 4, a HRRP radar target recognition method based on Inception-GAN is characterized in that: The method for extracting spatial features in step S2 is: First, the spatial distance between any two eigenvectors is calculated by calculating the Euclidean distance matrix, which is expressed as: Where D jk Represents X i The Euclidean distance between the feature vectors of the j-th and k-th nodes in , ||·||2 represents the Euclidean norm; The adjacency matrix is constructed based on the Euclidean distance matrix, and its expression is: Among them A i,jk Represents X i The connectivity relationship between the feature vectors of the j-th and k-th nodes in ; Then after obtaining the adjacency matrix, the adjacency matrix A i and the node feature matrix X are used as the input of the improved graph attention convolutional network to extract spatial structure features.
6. The HRRP radar target recognition method based on Inception-GAN according to claim 5 is characterized in that: In step S2, the adjacency matrix A i The expression of using the node feature matrix X as the input of the improved graph attention convolutional network to extract spatial structure features is: where Λ i ={j|A i,jk >0}, W is the linear transformation matrix; Among them, α ij (k) is the weight of each attention head k.
7. The HRRP radar target recognition method based on Inception-GAN according to claim 1 is characterized in that: In step S2, the extracted temporal features and spatial features are fused by a hybrid attention mechanism to obtain a fusion feature, which is expressed as: <h2 style=";text-align:left;direction:ltr">F<h2 style=";text-align:left;direction:ltr"> fusion <h2 style=";text-align:left;direction:ltr"> =softmax(W<h2 style=";text-align:left;direction:ltr"> f <h2 style=";text-align:left;direction:ltr"> ·[F1||F2]) Among them, W f is the learnable weight matrix.
8. The HRRP radar target recognition method based on Inception-GAN according to claim 1 is characterized in that: In step S3, the constructed Inception-GAN model is used as the feature extraction layer of the actor-critic network to construct the Inception-GAN deep reinforcement learning framework. During the construction process, the preprocessed HRRP sequence is used as the environment, and then the fused features obtained by the Inception-GAN model are input into the actor-critic network, so that the actor-critic network can interact with the fused features. At the same time, the classification features finally output by the construction framework can be continuously optimized by using the actor-critic network strategy update.
9. The HRRP radar target recognition method based on Inception-GAN according to claim 8 is characterized in that: In step S3, after constructing the Inception-GAN deep reinforcement learning framework, in order to enable the feature vector obtained after the next batch of data passes through the feature extraction layer to guide the agent to correctly select the next action, the convergence speed of the entire framework is improved by improving the hybrid loss function, which is expressed as: Among them, α, β, and γ are the weighting factors of the strategy function, value function, and loss function of the constructed Inception-GAN network, respectively.
10. The HRRP radar target recognition method based on Inception-GAN according to claim 1, characterized in that: In the step S4, the fused temporal features and spatial structural features are input into the constructed Inception-GAN deep reinforcement learning framework, and an optimized actor-critic network is obtained through the training of the intelligent agent, wherein the actor network outputs an optimized strategy, i.e., the probability distribution of selecting an action in each state; the critic network outputs an optimized state value estimate, which is used to evaluate the potential value of a given state; Under the framework of the agent and actor-critic network, the final classification result is output by selecting the best action in the optimal strategy, which is expressed as: C=arg max(π(A t |S t )) Among them, argmax is to select the category with the maximum probability in the probability distribution.