HRRP target identification method based on confrontation thought and integrated with clutter suppression

Abstract

The invention discloses an HRRP target recognition method integrated with clutter suppression based on an adversarial idea, and belongs to the field of radar target recognition, and the method comprises the following steps: S1, constructing an HRRP target data set which comprises clean HRRP target data and clutter-containing HRRP target data; s2, constructing a generator and a discriminator of a CN-LSGAN generation network to form a clutter suppression model; s3, training the clutter suppression model in an adversarial game mode; s4, performing clutter suppression on the input HRRP target signal by using the trained clutter suppression model to obtain HRRP target data; and S5, identifying the HRRP target data in an identification network containing an attention enhancement time encoder. According to the method, the clutter suppression problem is converted into the adversarial learning problem, the signal-to-noise ratio of the target HRRP is improved while the generative adversarial network is optimized, the clutter suppression effect is improved, and recognition is facilitated.

Description

Technical Field

[0001] The invention belongs to the field of radar target recognition, and in particular relates to a HRRP target recognition method with integrated clutter suppression based on a confrontation idea. Background Art

[0002] A high-resolution range profile (HRRP) is the vector sum of the target scattering point sub-echoes projected onto radar rays, acquired by a broadband radar signal. It reflects the distribution of the radar cross section (RCS) of the target scatterers along the radar line of sight (RLOS) at a specific radar viewing angle and contains important information such as target shape, orientation, and range. HRRP is one-dimensional data and, compared to synthetic aperture radar (SAR) and inverse synthetic aperture radar (ISAR) images, is easier to acquire and calculate.

[0003] The use of HRRP for radar target recognition is effective for identifying aerial targets such as aircraft and missiles. However, the difficulty of identifying sea surface targets using HRRP-based satellite radar is mainly reflected in the following two aspects: (1) The HRRP of sea surface targets of satellite radar is greatly interfered by sea clutter. The background sea clutter of the target is a strong interference factor. The waves and swells on the sea surface will produce complex electromagnetic scattering. The intensity and characteristics of its radar echo (sea clutter) will change dynamically with time, sea conditions (such as wind speed, wind direction, wave height) and other factors. At the same time, the statistical characteristics of sea clutter are difficult to accurately model. It usually does not obey a simple Gaussian distribution and may have characteristics such as long tails and non-stationarity. It is difficult to effectively suppress sea clutter using a single method. (2) There are various types of sea surface targets, including ships of different tonnages and purposes (such as cargo ships, oil tankers, warships, etc.), and there may also be targets such as buoys and small speedboats. The size, shape and material of different types of targets are different, and their HRRP characteristics will also vary greatly. At the same time, the HRRPs of some warships of the same type but different levels are similar. Therefore, when the existing radar target recognition method based on HRRP is applied to the spaceborne radar sea surface target recognition method, there are problems such as recognition mismatch and failure, and inaccurate classification. Summary of the Invention

[0004] The purpose of the present invention is to provide a HRRP target recognition method with integrated clutter suppression based on the countermeasure idea, so as to solve the problems of recognition mismatch and failure, inaccurate classification, etc. when the existing HRRP-based radar target recognition method is applied to the spaceborne radar sea surface target recognition method.

[0005] In order to achieve the above object, the technical solution of the present invention is as follows:

[0006] The present invention relates to a HRRP target recognition method with integrated clutter suppression based on a countermeasure concept, which comprises the following steps: S1. Construct a HRRP target dataset, including some clean HRRP target data and some HRRP target data containing clutter; S2. Construct the generator and discriminator of the CN-LSGAN generative network to form a clutter suppression model; S3. Input the HRRP target dataset and train the clutter suppression model using an adversarial game approach; S4 uses the trained clutter suppression model to suppress clutter on the input HRRP target signal to obtain HRRP target data; S5. Recognize HRRP target data using a recognition network containing an attention-enhanced temporal encoder.

[0007] Preferably, in the CN-LSGAN generation network constructed by S2, the losses of the generator and the discriminator are respectively expressed as: (4), in, G and D represent the generator and discriminator respectively, and Respectively represent the minimum value of the generator loss function and the minimum value of the discriminator loss function, and Respectively represent the loss function of the CN-LSGAN generation network and the loss function of the discriminant network, represents the gradient penalty term, and denote the regularization term, 、 and Represent the gradient penalty terms , regularization term as well as The coefficient of represents the classification loss, The coefficient representing the classification loss; The gradient penalty term is expressed as: (5), in, is the mixed input of the discriminator, is the expected operation of mixed input, is the discriminant output of the mixed input, For the discriminator D Mixed input The output gradient of is vectorized;

[0008] The regularization term and Respectively expressed as: (6), (7), in, x Clutter and x Clean They represent the HRRP target data input containing clutter and the clean HRRP target data input, E is the expectation operator; The classification loss is expressed as: (8), in, K Indicates the total number of HRRP target data input, t k Indicates the k The true category of the item, y k Indicates the k The predicted category of the item.

[0009] Preferably, the specific steps of training the clutter suppression model in S3 by using an adversarial game method include: S3.1. Fix the discriminator parameters, input the cluttered HRRP target data into the generator, continuously optimize the generator network parameters using the cluttered HRRP target data, and generate clutter-suppressed HRRP target data. S3.2. Fix the generator parameters, input the HRRP target data with suppressed clutter and the clean HRRP target data into the discriminator, continuously optimize the network parameters of the discriminator, and distinguish the HRRP target data; if the discriminator can distinguish the HRRP target data with suppressed clutter and the clean HRRP target data, return to S3.1 and perform generative training again; if the discriminator still cannot distinguish the HRRP target data with suppressed clutter and the clean HRRP target data after continuous updating, the training ends.

[0010] Preferably, the specific steps of S5 using a recognition network containing an attention-enhanced temporal encoder to identify HRRP target data are: S5.1. Generate HRRP sequences by continuously sliding a fixed-size sampling window; S5.2. Linearly map the HRRP sequence using the weight matrix to obtain the embedding vector of the HRRP target data. S5.3. Add position encoding to the embedding vector of HRRP target data; S5.4. For the embedded vector with position encoding, use N The layer-attention-enhanced temporal encoder computes attention scores in the temporal dimension; S5.5. Normalize the attention scores along the time dimension. S5.6. Input the normalized attention scores into the feedforward neural network for activation. S5.7. Perform layer-wise normalization on the activated attention scores. S5.8. After performing S5.5-S5.7 operations on the attention scores of each time dimension, perform vectorization operations and perform vectorization operations on the results. Soft The max operation obtains the target category.

[0011] Preferably, the specific steps of calculating the attention score in S5.4 are: S5.4.1. Calculate the row vectors of the key matrix, query matrix, and value matrix of each layer of the attention mechanism respectively. The calculation formula is: (9), in, , and Respectively represent n Key matrix of layer attention , query matrix Sum Matrix The row vector of Residual No. OK, , and represents the dimension transformation matrix, and , , ∈ R d×d , R represents a real matrix, d represents the matrix dimension, Layer normalization for computing the mean and variance of all layers for each input; S5.4.2. Calculate the n Layer attention score A n , the calculation formula is: (10), in, Soft max() represents the normalized exponential function, T Represents matrix transpose.

[0012] Preferably, the expression for the activation operation in S5.6 is: (11), in, FNN ( U n ) represents the normalized attention score U n Input to the feedforward neural network FNN In , ReLU is the activation function, W 1 and W 2 is the weight coefficient matrix, W 1∈ R d×d , W 2 ∈ R dr×d , R represents a real matrix, d represents the core feature dimension, dr express FNN The internal feature transformation dimension of .

[0013] Preferably, the expression of the S5.7 mid-level standardization operation is: (12), in, M n For the n The residual after the layer attention normalization operation, LN ()express LN function.

[0014] Preferably, in S5.8 The calculation formula is: (13), (14), Among them, y represents the target category, and its dimension is , s represents the n The residual after the layer attention normalization operation has the dimension , M n conductn The result of the vectorization operation of the last layer after layer encoding, Flatten Indicates flattening, and denote the weight matrix coefficients, and , , represents the dimension of the weight matrix coefficients, represents the dimension of the weight matrix coefficients, K Indicates the total number of input HRRP target data items.

[0015] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects: 1. This invention relates to a HRRP target recognition method with integrated clutter suppression based on adversarial thinking. This method targets HRRP of sea surface targets detected by spaceborne radar. A clutter suppression model is constructed using a CN-LSGAN generative network generator and discriminator. This model is trained using an adversarial game approach. The clutter suppression model employs an adversarial integrated clutter suppression method to suppress target clutter. This method transforms the clutter suppression problem into an adversarial learning problem, optimizing the generative adversarial network while improving the signal-to-noise ratio of the target HRRP and achieving clutter suppression.

[0016] 2. The HRRP target recognition method with integrated clutter suppression based on adversarial thinking involved in the present invention adopts a recognition network containing an attention-enhanced temporal encoder to recognize HRRP target data. The recognition network of the attention-enhanced temporal encoder starts from solving the long-term and global dependency acquisition problem of HRRP. The recognition module adopts an attention-enhanced temporal encoder with parallelization and global sequential feature extraction, and emphasizes important input data by calculating the attention score, thereby weighting the feature vector for easy recognition.

[0017] 3. The HRRP target recognition method with integrated clutter suppression based on adversarial thinking proposed by the present invention integrates clutter suppression and HRRP data recognition. Compared with the traditional separation method (which usually treats clutter suppression and target recognition as two independent processes and handles them separately), it avoids information loss, improves system efficiency, and achieves end-to-end optimization through integrated learning technology, thereby improving the overall performance of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Schematic diagram of the HRRP target recognition model with integrated clutter suppression based on the countermeasure concept of the present invention; Figure 2 Schematic diagram of the CN-LSGAN clutter suppression module of the present invention; Figure 3This is a schematic diagram of the idea of ​​integrating clutter suppression and identification countermeasures in the present invention; Figure 4 Schematic diagram of the recognition network module of the attention-enhanced temporal encoder of the present invention. DETAILED DESCRIPTION

[0019] In order to further understand the content of the present invention, the present invention is described in detail with reference to the examples. The following examples are used to illustrate the present invention but are not used to limit the scope of the present invention.

[0020] Refer to the attached Figure 1 As shown, the present invention relates to a HRRP target recognition method with integrated clutter suppression based on the countermeasure idea, which comprises the following steps: S1. Construct an HRRP target dataset, including some clean HRRP target data and some HRRP target data containing clutter.

[0021] S2. Build the generator and discriminator of the CN-LSGAN generative network to form a clutter suppression model: The CN-LSGAN generative network constructed in S2 is represented by the generator constraint term and the discriminator constraint Constrained CNGAN network.

[0022] The generator constraint and the discriminator constraint are respectively expressed as: (1), (2), Among them, the discriminator constraint Helps solve model collapse, generator constraints The degree of freedom of G can be reduced, thereby more efficiently improving the signal-to-noise ratio of the generated HRRP target data. G and D represent the generator and discriminator respectively, x is the HRRP target data of the input generator, G ( z ) is the HRRP target data generated by the generator, E [·] is the expectation operator, z For input to the generator G The HRRP target data containing clutter in P z is the probability distribution of HRRP containing the target, is the mixed input of the discriminator, which is the data obtained by mixing the HRRP target data generated by the generator with the clean HRRP data. It is the discriminant output of mixed input, used to judge the mixed input Is it clean HRRP data or generated HRRP data? is the probability distribution of the mixed input.

[0023] The CN-LSGAN generation network is expressed as: (3), in, V The objective function value of the CN-LSGAN generation network is an indicator of network performance obtained by comprehensively considering the generator constraints and the discriminator constraints. and Represents the adjustment weight, which is used to adjust the generator constraint and the discriminator constraint respectively. P z ( x ) is the generated data distribution output by the generator, P data ( x ) is the clean HRRP target data distribution; In the above CN-LSGAN generation network, the losses of the generator and discriminator are expressed as: (4), in, G and D represent the generator and discriminator respectively, and Respectively represent the minimum value of the generator loss function and the minimum value of the discriminator loss function, and Respectively represent the loss function of the CN-LSGAN generation network and the loss function of the discriminant network, represents the gradient penalty term, and denote the regularization term, 、 and Represent the gradient penalty terms , regularization term as well as The coefficient of represents the classification loss, The coefficient representing the classification loss; The gradient penalty term is expressed as: (5), in, is the mixed input of the discriminator, is the expected operation of mixed input, is the discriminant output of the mixed input, For the discriminator D Mixed input The output gradient of is vectorized; The regularization term and Respectively expressed as: (6), (7), in, x Clutter and x Clean They represent the HRRP target data input containing clutter and the clean HRRP target data input, E is the expectation operator; The classification loss is expressed as: (8), in, K Indicates the total number of HRRP target data input, t k Indicates the k The true category of the item, y k Indicates the k The predicted category of the item.

[0024] The CN-LSGAN generative network is optimized from a naive GAN. The generator G and discriminator D of the naive GAN improve their respective capabilities through adversarial learning, ultimately generating "real" data. The learning process of the generator G and discriminator D is expressed as the following objective function: (15); Naive GAN is prone to the gradient vanishing problem during model learning. To overcome this problem, we refer to the Least-Squares GAN method and transform Equation (14) into the objective function in the sense of least squares, as shown in Equation (15): (16); Since the naive GAN has a large output freedom problem, it often generates unreasonable samples that can deceive the discriminator D. To solve this problem, a regression function constraint term as shown in formula (2) is added to the target optimization function of the generator G to minimize the mean square distance between the generated sample and the real signal, thereby reducing the output freedom of the generator G and enabling the improved naive GAN to solve the data generation problem. Therefore, formula (15) can be written as formula (16): (17); The naive GAN with the above constraints is called constrained naive GAN (CN-GAN).

[0025] Adding generator constraints After that, because the real data is visible to the generator G, the training of the generator G actually changes from unsupervised learning to supervised learning. In formula (16), the loss function of the generator G consists of two parts: the first part is the adversarial loss function of the generator G, and the second part is the sample distortion loss function. If the adversarial loss function is removed (the learning of the generator G does not depend on the discriminator D), the learning of the generator G is similar to the current mainstream deep learning based on a single deep neural network. Although the single deep neural network has good nonlinear learning ability, when the signal-to-interference-clutter ratio is low, due to the influence of strong interference and clutter, the nonlinear learning ability will also degenerate. Therefore, it can be seen that CN-GAN guides the generator training through the discriminator, which can make up for the design deficiencies of the loss function of the single deep neural network and solve the performance degradation problem caused by low signal-to-interference-clutter ratio, and obtain better training results.

[0026] S3. Input the HRRP target data set and train the clutter suppression model using adversarial game method; the specific steps are as follows: Figure 2 and Figure 3 As shown, including: S3.1. Fix the discriminator parameters, input the cluttered HRRP target data into the generator, continuously optimize the generator network parameters using the cluttered HRRP target data, and generate clutter-suppressed HRRP target data. S3.2. Fix the generator parameters, input the HRRP target data with suppressed clutter and the clean HRRP target data into the discriminator, continuously optimize the network parameters of the discriminator, and distinguish the HRRP target data; if the discriminator can distinguish the HRRP target data with suppressed clutter and the clean HRRP target data, return to S3.1 and perform generative training again; if the discriminator still cannot distinguish the HRRP target data with suppressed clutter and the clean HRRP target data after continuous updating, it means that Nash equilibrium is reached and the training ends.

[0027] S4. Use the trained clutter suppression model to suppress clutter on the input HRRP target signal to obtain HRRP target data.

[0028] S5. Use the recognition network with attention-enhanced temporal encoder to identify HRRP target data. The specific steps are as follows: Figure 4 As shown, S5.1. Generate HRRP sequences by continuously sliding a fixed-size sampling window; S5.2. Linearly map the HRRP sequence using the weight matrix to obtain the embedding vector of the HRRP target data. S5.3. Add position encoding to the embedding vector of HRRP target data; S5.4. For the embedded vector with position encoding, use N The layer-attention-enhanced temporal encoder calculates the attention score in the time dimension. The specific steps for calculating the attention score are: S5.4.1. Calculate the row vectors of the key matrix, query matrix, and value matrix of each layer of the attention mechanism respectively. The calculation formula is: (9), in, , and Respectively represent n Key matrix of layer attention , query matrix Sum Matrix The row vector of Residual No. OK, , and represents the dimension transformation matrix, and , , ∈ R d×d , R represents a real matrix, d represents the matrix dimension, Layer normalization for computing the mean and variance of all layers for each input; S5.4.2. Calculate the n Layer attention score A n , the calculation formula is: (10), in, Soft max() represents the normalized exponential function, T Represents matrix transpose.

[0029] S5.5. Normalize the attention scores along the time dimension. S5.6. Input the normalized attention score into the feedforward neural network for activation operation. The expression of the activation operation is: (11), in, FNN ( U n ) represents the normalized attention score U n Input to the feedforward neural network FNN In , ReLU is the activation function, W 1 and W 2 is the weight coefficient matrix, W 1∈ R d×d , W 2 ∈ R dr×d , R represents a real matrix, d represents the core feature dimension, dr express FNN The internal feature transformation dimension of S5.7. Perform layer normalization on the activated attention score. The expression of layer normalization is: (12), in, M n For the n The residual after the layer attention normalization operation, LN ()express LN function; S5.8. After performing S5.5-S5.7 operations on the attention scores of each time dimension, perform vectorization operations and perform vectorization operations on the results. Soft The max operation gets the target category, The calculation formula is: (13), (14), Among them, y represents the target category, and its dimension is , s represents the n The residual after the layer attention normalization operation has the dimension , M n conduct n The result of the vectorization operation of the last layer after layer encoding, Flatten Indicates flattening, and denote the weight matrix coefficients, and , , represents the dimension of the weight matrix coefficients, represents the dimension of the weight matrix coefficients, K Indicates the total number of input HRRP target data items.

[0030] The present invention has been described in detail above with reference to the embodiments. However, the contents described are only preferred embodiments of the present invention and should not be considered as limiting the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A HRRP target recognition method with integrated clutter suppression based on adversarial thinking, characterized by: It includes the following steps: S1. Construct a HRRP target dataset, including some clean HRRP target data and some HRRP target data containing clutter; S2. Construct the generator and discriminator of the CN-LSGAN generative network to form a clutter suppression model; S3. Input the HRRP target dataset and train the clutter suppression model using an adversarial game approach; S4 uses the trained clutter suppression model to suppress clutter on the input HRRP target signal to obtain HRRP target data; S5. Recognize HRRP target data using a recognition network containing an attention-enhanced temporal encoder.

2. The HRRP target recognition method with integrated clutter suppression based on the countermeasure concept according to claim 1 is characterized in that: In the CN-LSGAN generation network constructed by S2, the losses of the generator and discriminator are expressed as: （4）， in, G and D represent the generator and discriminator respectively, and Respectively represent the minimum value of the generator loss function and the minimum value of the discriminator loss function, and Respectively represent the loss function of the CN-LSGAN generation network and the loss function of the discriminant network, represents the gradient penalty term, and denote the regularization term, 、 and Represent the gradient penalty terms , regularization term as well as The coefficient of represents the classification loss, The coefficient representing the classification loss; The gradient penalty term is expressed as: （5）， in, is the mixed input of the discriminator, is the expected operation of mixed input, is the discriminant output of the mixed input, For the discriminator D Mixed input The output gradient of is vectorized; The regularization term and Respectively expressed as: （6）， （7）， in, x Clutter and x Clean They represent the HRRP target data input containing clutter and the clean HRRP target data input, E is the expectation operator; The classification loss is expressed as: （8）， in, K Indicates the total number of HRRP target data input, t k Indicates the k The true category of the item, y k Indicates the k The predicted category of the item.

3. The HRRP target recognition method with integrated clutter suppression based on the countermeasure concept according to claim 1 is characterized in that: The specific steps of S3 using the adversarial game method to train the clutter suppression model include: S3.

1. Fix the discriminator parameters, input the cluttered HRRP target data into the generator, continuously optimize the generator network parameters using the cluttered HRRP target data, and generate clutter-suppressed HRRP target data. S3.

2. Fix the generator parameters, input the HRRP target data with suppressed clutter and the clean HRRP target data into the discriminator, continuously optimize the network parameters of the discriminator, and distinguish the HRRP target data; if the discriminator can distinguish the HRRP target data with suppressed clutter and the clean HRRP target data, return to S3.1 and perform generative training again; if the discriminator still cannot distinguish the HRRP target data with suppressed clutter and the clean HRRP target data after continuous updating, the training ends.

4. The HRRP target recognition method with integrated clutter suppression based on the countermeasure concept according to claim 1 is characterized in that: The specific steps of S5 using the recognition network containing the attention-enhanced temporal encoder to identify the HRRP target data are as follows: S5.

1. Generate HRRP sequences by continuously sliding a fixed-size sampling window; S5.

2. Linearly map the HRRP sequence using the weight matrix to obtain the embedding vector of the HRRP target data. S5.

3. Add position encoding to the embedding vector of HRRP target data; S5.

4. For the embedded vector with position encoding, use N The layer-attention-enhanced temporal encoder computes attention scores in the temporal dimension; S5.

5. Normalize the attention scores along the time dimension. S5.

6. Input the normalized attention scores into the feedforward neural network for activation. S5.

7. Perform layer-wise normalization on the activated attention scores. S5.

8. After performing S5.5-S5.7 operations on the attention scores of each time dimension, perform vectorization operations and perform vectorization operations on the results. Soft The max operation obtains the target category.

5. The HRRP target recognition method with integrated clutter suppression based on the countermeasure concept according to claim 4 is characterized in that: The specific steps of calculating the attention score in S5.4 are as follows: S5.4.

1. Calculate the row vectors of the key matrix, query matrix, and value matrix of each layer of the attention mechanism respectively. The calculation formula is: （9）， in, , and Respectively represent n Key matrix of layer attention , query matrix Sum Matrix The row vector of Residual No. OK, , and represents the dimension transformation matrix, and , , ∈ R d×d , R represents a real matrix, d represents the matrix dimension, Layer normalization for computing the mean and variance of all layers for each input; S5.4.

2. Calculate the n Layer attention score A n , the calculation formula is: （10）， in, Soft max() represents the normalized exponential function, T Represents matrix transpose.

6. The HRRP target recognition method with integrated clutter suppression based on the countermeasure concept according to claim 5 is characterized in that: The expression of the activation operation in S5.6 is: （11）， in, FNN ( U n ) represents the normalized attention score U n Input to the feedforward neural network FNN In , ReLU is the activation function, W 1 and W 2 is the weight coefficient matrix, W 1∈ R d×d , W 2 ∈ R dr×d , R represents a real matrix, d represents the core feature dimension, dr express FNN The internal feature transformation dimension of .

7. The HRRP target recognition method with integrated clutter suppression based on the countermeasure concept according to claim 6 is characterized in that: The expression of the mid-level normalization operation in S5.7 is: （12）， in, M n For the n The residual after the layer attention normalization operation, LN ()express LN function.

8. The HRRP target recognition method with integrated clutter suppression based on the countermeasure concept according to claim 7 is characterized in that: S5.8 The calculation formula is: （13）， （14）， Among them, y represents the target category, and its dimension is , s represents the n The residual after the layer attention normalization operation has the dimension , M n conduct n The result of the vectorization operation of the last layer after layer encoding, Flatten Indicates flattening, and denote the weight matrix coefficients, and , , represents the dimension of the weight matrix coefficients, represents the dimension of the weight matrix coefficients, K Indicates the total number of input HRRP target data items.

Images