Hrrp target recognition method based on integrated clutter suppression of adversarial thought

Abstract

The application discloses an HRRP target recognition method based on integrated clutter suppression of an adversarial idea and belongs to the field of radar target recognition, which comprises the following steps: S1, constructing an HRRP target data set, including clean HRRP target data and HRRP target data containing clutter; 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, using the trained clutter suppression model to perform clutter suppression on input HRRP target signals to obtain HRRP target data; and S5, using an identification network containing an attention-enhanced time encoder to identify the HRRP target data. The application converts the clutter suppression problem into an adversarial learning problem, optimizes the generative adversarial network, improves the signal-to-noise ratio of the target HRRP, improves the effect of clutter suppression, and facilitates identification.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of radar target recognition, and particularly relates to an HRRP target recognition method based on integrated clutter suppression of an adversarial idea. BACKGROUND

[0002] High Resolution Range Profile (HRRP) is a vector sum of target scattering point subechoes on radar rays obtained by wideband radar signal, which reflects the distribution of radar scattering cross section (RCS) of a scattering body on a target along a radar line of sight (RLOS) under a specific radar view, and contains important information such as target shape, orientation and distance. HRRP is one-dimensional data, and compared with a synthetic aperture radar (SAR) image and an inverse synthetic aperture radar (ISAR) image, HRRP has the characteristics of easy acquisition and calculation.

[0003] Radar target recognition using HRRP has good recognition effect on aerial targets such as aircraft and missiles, but the difficulties of sea surface target recognition based on HRRP of a spaceborne radar mainly lie in the following two aspects: (1) HRRP of a spaceborne radar sea surface target is greatly interfered by sea clutter, and the background sea clutter of a target is a very strong interference factor. Electromagnetic scattering of sea waves and swell of a sea surface produces complex electromagnetic scattering, and the intensity and characteristics of radar echoes (sea clutter) dynamically change with factors such as time, sea conditions (such as wind speed, wind direction and wave height). Meanwhile, the statistical characteristics of sea clutter are difficult to accurately model. It usually does not conform to simple Gaussian distribution, and may have long tailing, non-stationary and other characteristics. It is difficult to effectively suppress sea clutter using a single means. (2) Sea surface targets are various, including various tonnage and purpose ships (such as cargo ships, oil tankers and warships), and there may be buoys and small speedboats. The sizes, shapes and materials of different types of targets are different, and the HRRP characteristics are also very different. Meanwhile, the HRRPs of some military ships of the same type but different levels have similarities. Therefore, when the existing radar target recognition method based on HRRP is applied to sea surface target recognition of a spaceborne radar, there are problems such as recognition mismatch and failure, and inability to accurately classify. SUMMARY

[0004] The application aims to provide an HRRP target recognition method based on integrated clutter suppression in an adversarial idea, so as to solve the problems of recognition mismatch and failure, and inability to accurately classify, etc. when the existing HRRP-based radar target recognition method is applied to the sea surface target recognition method of a spaceborne radar.

[0005] In order to achieve the above-mentioned purpose, the technical scheme of the application is as follows:

[0006] The application relates to an HRRP target recognition method based on integrated clutter suppression in an adversarial idea, which comprises the following steps:

[0007] S1. Constructing an HRRP target data set, including part of clean HRRP target data and part of HRRP target data containing clutter;

[0008] S2. Constructing a generator and a discriminator of a CN-LSGAN generation network to form a clutter suppression model;

[0009] S3. Inputting the HRRP target data set, and training the clutter suppression model in an adversarial game mode;

[0010] S4. Using the trained clutter suppression model to suppress the clutter of the input HRRP target signal to obtain HRRP target data;

[0011] S5. Recognizing the HRRP target data in a recognition network containing an attention-enhanced time encoder.

[0012] Preferably, in the CN-LSGAN generation network constructed in S2, the loss of the generator and the loss of the discriminator are respectively represented as:

[0013] (4),

[0014] wherein, G and D represent the generator and the discriminator respectively, and represent the minimum value of the loss function of the generator and the minimum value of the loss function of the discriminator respectively, and represent the loss function of the CN-LSGAN generation network and the loss function of the discriminator network respectively, represents a gradient penalty term, and represent regularization terms respectively, , and represent the coefficients of the gradient penalty term , the regularization term and respectively, a classification loss, a coefficient of the classification loss;

[0015] The gradient penalty term is expressed as:

[0016] (5),

[0017] wherein, is a mixed input of the discriminator, is an expectation operation of the mixed input, is a discriminant output of the mixed input, is a discriminator D The output gradient of the mixed input is calculated by vectorization;

[0018] The regularization term and are respectively expressed as:

[0019] (6),

[0020] (7),

[0021] wherein, x Clutter and x Clean respectively represent the clutter-containing HRRP target data input and the clean HRRP target data input, E is an expectation operator;

[0022] The classification loss is expressed as:

[0023] (8),

[0024] wherein, K represents the total number of input HRRP target data, t k represents the true class of the k th item, y k represents the predicted class of the k th item.

[0025] Preferably, the specific steps of training the clutter suppression model by the S3 in the manner of an adversarial game include:

[0026] S3.1. Fix the discriminator parameters, input the clutter-containing HRRP target data into the generator, and constantly optimize the network parameters of the generator with the clutter-containing HRRP target data to generate the HRRP target data suppressed by clutter;

[0027] S3.2. Fix the generator parameters, input the HRRP target data with suppressed clutter and the clean HRRP target data into the discriminator, constantly 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 to generate training again; if the constantly updated discriminator still cannot distinguish the HRRP target data with suppressed clutter and the clean HRRP target data, the training is ended.

[0028] Preferably, the specific steps of identifying the HRRP target data in the identification network containing the attention-enhanced time encoder in the S5 are as follows:

[0029] S5.1. Generate the HRRP sequence by continuously sliding a sampling window with a fixed size;

[0030] S5.2. Linearly map the HRRP sequence by a weight matrix to obtain the embedding vector of the HRRP target data;

[0031] S5.3. Add position encoding to the embedding vector of the HRRP target data;

[0032] S5.4. For the embedding vector with the added position encoding, the attention-enhanced time encoder is used to calculate the attention score; N The layer attention-enhanced time encoder calculates the attention score in the time dimension;

[0033] S5.5. Normalize the attention score in the time dimension;

[0034] S5.6. Input the normalized attention score into the feedforward neural network for activation operation;

[0035] S5.7. Perform layer normalization operation on the activated attention score;

[0036] S5.8. After the operations of S5.5-S5.7 on the attention score in each time dimension, perform vectorization operation, and perform max operation on the vectorization operation result to obtain the target category. Soft

[0037] Preferably, the specific steps of calculating the attention score in the S5.4 are as follows:

[0038] S5.4.1. Calculate the row vectors of the key matrix, the query matrix and the value matrix of each layer attention mechanism respectively, and the calculation formula is as follows:

[0039] (9),

[0040] wherein, , and​ respectively represent the first n key matrix of the layer attention , the query matrix and the value matrix , represent the residual of the first row, , and represent the dimension transformation matrix, and , , ∈ R d×d , R represent the real number matrix, d represent the matrix dimension, layer normalization is used to calculate the mean and variance of all layers of each input;

[0041] S5.4.2. Calculate the first n layer attention score A n , the calculation formula is:

[0042] (10),

[0043] wherein, Soft max() represents the normalized exponential function, T represents the matrix transpose.

[0044] Preferably, the expression of the activation operation in S5.6 is:

[0045] (11),

[0046] wherein, FNN ( U n ) represents the attention score U n input into the feedforward neural network FNN , ReLU is the activation function, W 1 and W 2 are weight coefficient matrices, W 1∈ R d×d , W 2 ∈ R dr×d , R represent the real number matrix, d represent the core feature dimension, dr represent the internal feature transformation dimension of FNN .

[0047] Preferably, the expression of the layer standardization operation in S5.7 is:

[0048] (12),

[0049] wherein, M n is the first n layer attention standardization operation after the residual, LN () represents LN a function.

[0050] Preferably, the operation formula of the operation in S5.8 is:

[0051] (13),

[0052] (14), wherein, y represents a target category, and the dimension of y is

[0053] , s represents the residual after the first layer attention standardization operation, and the dimension of s is n , , M n performing n vectorization operation on the last layer after the layer encoding, Flatten represents flattening, and respectively represent weight matrix coefficients, and , , represents the dimension of the weight matrix coefficient, represents the dimension of the weight matrix coefficient, K represents the total number of items of the input HRRP target data.

[0054] Compared with the prior art, the technical scheme provided by the present application has the following beneficial effects:

[0055] 1. The HRRP target recognition method based on the integrated clutter suppression of the adversarial idea provided by the present application is aimed at the HRRP of the sea surface target of the spaceborne radar, constructs a clutter suppression model composed of the generator and the discriminator of the CN-LSGAN generation network, and trains the clutter suppression model in the manner of the adversarial game. The clutter suppression model uses the integrated clutter suppression method based on the adversarial idea to suppress the clutter of the target. In the integrated clutter suppression method based on the adversarial idea, the clutter suppression problem is converted into the problem of adversarial learning, optimizes the generative adversarial network, improves the signal-to-noise ratio of the target HRRP, and realizes the effect of clutter suppression.

[0056] 2.The HRRP target recognition method based on the integrated clutter suppression and the adversarial idea according to the application adopts the recognition network containing the attention-enhanced time encoder to recognize the HRRP target data, the recognition network of the attention-enhanced time encoder is started from solving the long-term and global dependence acquisition problem of the HRRP, the recognition module adopts the attention-enhanced time encoder with parallelization and global sequential feature extraction, important input data is emphasized by calculating the attention score, so that the feature vector is weighted to facilitate recognition.

[0057] 3.The HRRP target recognition method based on the integrated clutter suppression and the adversarial idea according to the application integrates the clutter suppression and the recognition of the HRRP data, compared with the traditional separation method (generally, the clutter suppression and the target recognition are regarded as two independent processes and are processed respectively), information loss is avoided, the system efficiency is improved, end-to-end optimization is realized through the integrated learning technology, and the overall performance of the system is improved. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 It is a HRRP target recognition model based on the integrated clutter suppression and the adversarial idea according to the application;

[0059] Figure 2 It is a CN-LSGAN clutter suppression module according to the application;

[0060] Figure 3 It is an integrated clutter suppression and recognition adversarial idea according to the application;

[0061] Figure 4 It is an attention-enhanced time encoder recognition network module according to the application. DETAILED DESCRIPTION

[0062] In order to further understand the content of the application, the application will be described in detail in combination with the embodiments, and the following embodiments are used to illustrate the application, but not to limit the scope of the application.

[0063] With reference to the accompanying Figure 1 drawings, the application relates to an HRRP target recognition method based on the integrated clutter suppression and the adversarial idea, which comprises the following steps:

[0064] S1.Building an HRRP target data set, including part of clean HRRP target data and part of HRRP target data containing clutter.

[0065] S2.Building a generator and a discriminator of a CN-LSGAN generation network to form a clutter suppression model:

[0066] The CN-LSGAN generation network built in S2 is represented by a generator constraint term and a discriminator constraint term Constrained CNGAN network.

[0067] The generator constraint and the discriminator constraint are respectively expressed as:

[0068] (1),

[0069] (2),

[0070] 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.

[0071] The CN-LSGAN generation network is expressed as:

[0072] (3),

[0073] 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;

[0074] In the CN-LSGAN generation network, the loss of the generator and the discriminator is respectively represented as:

[0075] (4),

[0076] wherein, G and D respectively represent the generator and the discriminator, 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 discriminator network, represents a gradient penalty term, and respectively represent a regularization term, , and respectively represent the coefficient of the gradient penalty term , the regularization term and , represents a classification loss, represents the coefficient of the classification loss;

[0077] The gradient penalty term is represented as:

[0078] (5),

[0079] wherein, is the mixed input of the discriminator, is the expected operation of the mixed input, is the discriminant output of the mixed input, is the output gradient vectorization calculation of the discriminator D to the mixed input ;

[0080] The regularization term and are respectively represented as:

[0081] (6),

[0082] (7),

[0083] wherein, x Clutter and x Clean respectively represent the HRRP target data input containing clutter and the clean HRRP target data input, EDesired operator

[0084] The classification loss is represented as:

[0085] (8),

[0086] Wherein, K N represents the total number of input HRRP target data, t k Y represents the true class of the i-th item, k y k Y represents the predicted class of the i-th item. k

[0087] The CN-LSGAN generation network is optimized from the naive GAN, and the generator G and the discriminator D of the naive GAN improve their respective abilities through confrontation, and finally generate "real" data. The learning process of the generator G and the discriminator D is represented as the following objective function:

[0088] (15);

[0089] The naive GAN is prone to gradient disappearance problem during model learning. In order to overcome this problem, the method of Least-Squares GAN is used to convert formula (14) into the objective function in the least square sense, as shown in formula (15):

[0090] (16);

[0091] Because the naive GAN has the problem of large output degree of freedom, it often generates unreasonable samples that can deceive the discriminator D. In order to solve this problem, a regression function constraint term as shown in formula (2) is added to the objective optimization function of the generator G, so that the mean square distance between the generated samples and the real signal is minimized, thereby reducing the output degree of freedom of the generator G. The improved naive GAN can solve the data generation problem;

[0092] Therefore, formula (15) can be written as formula (16):

[0093] (17);

[0094] The naive GAN with the above constraint term is called constrained naive GAN (CN-GAN).

[0095] Adding generator constraint term ​​After that, because the real data is visible to the generator G, the training of the generator G is actually converted from unsupervised learning to supervised learning, and the loss function of the generator G in formula (16) is composed of two parts, the former part is the adversarial loss function of the generator G, and the latter 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, but when the low signal-to-interference ratio, due to the influence of strong interference and clutter, the nonlinear learning ability will also degenerate. Therefore, the CN-GAN can make up for the design deficiency of the loss function of the single deep neural network and solve the performance degradation problem caused by the low signal-to-interference ratio by guiding the generator training through the discriminator, and obtain better training effect.

[0096] S3. Input the HRRP target data set, and train the clutter suppression model in an adversarial game manner; the specific steps are as shown in Figure 2 and Figure 3 , and include:

[0097] S3.1. Fix the discriminator parameter, input the HRRP target data containing clutter into the generator, constantly optimize the network parameters of the generator with the HRRP target data containing clutter, and generate the HRRP target data with suppressed clutter;

[0098] S3.2. Fix the generator parameter, input the HRRP target data with suppressed clutter and the clean HRRP target data into the discriminator, constantly 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 for generation training; if the constantly updated discriminator still cannot distinguish the HRRP target data with suppressed clutter and the clean HRRP target data, i.e. Nash equilibrium is reached, the training is completed.

[0099] S4. Use the trained clutter suppression model to suppress the input HRRP target signal, and obtain the HRRP target data.

[0100] S5. Use the recognition network containing the attention-enhanced time encoder to recognize the HRRP target data, and the specific steps are as shown in Figure 4 , and include:

[0101] S5.1. Generate the HRRP sequence by continuously sliding a sampling window of a fixed size;

[0102] S5.2. Linearly map the HRRP sequence through a weight matrix to obtain the embedding vector of the HRRP target data;

[0103] S5.3. Add the position code to the embedding vector of the HRRP target data;

[0104] S5.4. For the embedding vector added with the position code, use N The layer attention enhanced time encoder calculates the attention score in the time dimension, and the specific steps of calculating the attention score are as follows:

[0105] S5.4.1. Calculate the row vectors of the key matrix, the query matrix and the value matrix of each layer attention mechanism respectively, and the calculation formula is as follows:

[0106] (9),

[0107] wherein, , and represent the row vectors of the key matrix n , the query matrix and the value matrix of the first layer attention, represent the first row of the residual , , and represent the dimension transformation matrix, and , , ∈ R d×d , R represent the real matrix, d represent the matrix dimension, layer normalization for calculating the mean and variance of all layers of each input;

[0108] S5.4.2. Calculate the first n layer attention score A n , and the calculation formula is as follows:

[0109] (10),

[0110] wherein, Soft max() represents the normalized exponential function, T represents the matrix transpose.

[0111] S5.5. Normalize the attention score in the time dimension;

[0112] S5.6. Input the normalized attention score into the feedforward neural network for activation operation, and the expression of the activation operation is as follows:

[0113] (11),

[0114] wherein, FNN ( U n ) denotes the normalized attention score U n input into the feedforward neural network FNN , ReLU is an activation function, W 1 and W 2 are weight coefficient matrices, W 1∈ R d×d , W 2 ∈ R dr×d , R denotes a real matrix, d denotes a core feature dimension, dr denotes FNN an internal feature transformation dimension;

[0115] S5.7. Perform layer normalization operation on the activated attention score, and the expression of the layer normalization operation is:

[0116] (12),

[0117] wherein, M n is the residual after the n layer attention normalization operation, LN () denotes LN a function;

[0118] S5.8. After performing the operations of S5.5-S5.7 on the attention score in each time dimension, perform vectorization operation, and perform Soft max operation on the vectorization operation result to obtain the target category, The operation formula of the operation is:

[0119] (13),

[0120] (14),

[0121] wherein, y denotes the target category, and the dimension of y is , s denotes the residual after the n layer attention normalization operation, and the dimension of s is , M n perform vectorization operation on the last layer after n layer encoding, Flatten denotes flattening, and respectively represent the weight matrix coefficients, and , , represents the dimension of the weight matrix coefficients, represents the dimension of the weight matrix coefficients, K represents the total number of items of the input HRRP target data.

[0122] The application has been described in detail with reference to the embodiments. However, the above description is only the preferred embodiments of the application, and should not be considered as limiting the scope of the application. Any equivalent changes and improvements made according to the scope of the application should still fall within the scope of the patent.

Claims

1. A HRRP target recognition method based on integrated clutter suppression of adversarial thinking, characterized in that: It comprises the following steps: S1. Constructing an HRRP target data set, including partially clean HRRP target data and partially cluttered HRRP target data; S2. Constructing a generator and a discriminator of a CN-LSGAN generation network to form a clutter suppression model, wherein the loss of the generator and the discriminator in the constructed CN-LSGAN generation network is respectively represented as: （4）， wherein, G and D denote the generator and the discriminator, respectively, and denote taking the minimum value of the generator loss function and the minimum value of the discriminator loss function, respectively, and denote the loss function of the CN-LSGAN generation network and the loss function of the discrimination network, respectively, denotes a gradient penalty term, and denote regularization terms, respectively, , and denote the coefficient of the gradient penalty term , the regularization term , and , respectively, denotes a classification loss, denotes the coefficient of the classification loss; The gradient penalty term is represented as: （5）， wherein, is a mixed input to a discriminator, is a desired operation on the mixed input, is a discriminative output of the mixed input, is a discriminator D on the mixed input is output gradient vectorization computation; The regularization term and are respectively expressed as: （6）， （7）， wherein, x Clutter and x Clean respectively represent the HRRP target data input containing clutter and the clean HRRP target data input, E is a desired operator; The classification loss is represented as: （8）， wherein, K represents the total number of input HRRP target data, t k represents the true class of the k th item, y k represents the predicted class of the k th item; S3. Inputting the HRRP target data set and training the clutter suppression model in an adversarial game manner, comprising the following steps: S3.

1. Fixing the parameters of the discriminator, inputting the cluttered HRRP target data into the generator, constantly optimizing the network parameters of the generator with the cluttered HRRP target data, and generating HRRP target data with suppressed clutter; S3.

2. Fixing the parameters of the generator, inputting the HRRP target data with suppressed clutter and the clean HRRP target data into the discriminator, constantly optimizing the network parameters of the discriminator, and distinguishing the HRRP target data; if the discriminator can distinguish the HRRP target data with suppressed clutter and the clean HRRP target data, returning to S3.1 for generation training; if the constantly updated discriminator still cannot distinguish the HRRP target data with suppressed clutter and the clean HRRP target data, the training is ended; S4. Using the trained clutter suppression model to suppress the clutter of the input HRRP target signal to obtain HRRP target data; S5. Using an identification network with an attention-enhanced time encoder to identify the HRRP target data, comprising the following steps: S5.

1. Generating an HRRP sequence by continuously sliding a sampling window of a fixed size; S5.

2. Linearly mapping the HRRP sequence through a weight matrix to obtain an embedding vector of the HRRP target data; S5.

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

4. For the embedded vector of the join location coding, employ N The layer attention enhanced temporal encoder computes attention scores over the temporal dimension. S5.

5. Normalizing the attention score in the time dimension; S5.

6. Inputting the normalized attention score into a feedforward neural network for activation operation; S5.

7. Performing layer normalization operation on the activated attention score; S5.

8. After the operations of S5.5-S5.7 are performed on the attention scores of each time dimension, vectorization operations are performed on the attention scores of each time dimension, and the vectorization operation results are Soft max operation to obtain the target category.

2. The HRRP target recognition method based on integrated clutter suppression according to the thought of confrontation according to claim 1, characterized in that: The specific steps of calculating the attention score in S5.4 are as follows: S5.4.

1. Calculating the row vectors of the key matrix, the query matrix and the value matrix of each layer of attention mechanism respectively, and the calculation formula is as follows: （9）， wherein, , and denote the key matrix n layer attention , the query matrix and the value matrix , denote the row vector of the residual , row, , and denote the dimension transformation matrix, and , , ∈ R d×d , R denote the real matrix, d denote the matrix dimension, layer normalization for calculating the mean and variance of all layers for each input; S5.4.

2. Computing the first n Layer attention scores A n , the formula is: （10）， wherein Soft max() denotes the normalized exponential function, T denotes the matrix transpose.

3. The HRRP target recognition method based on integrated clutter suppression of the adversarial thought according to claim 2, characterized in that: The expression of the activation operation in S5.6 is as follows: （11）， wherein, FNN U n denotes the normalized attention score U n input to the feedforward neural network FNN ReLU is an activation function, W 1 and W 2 are weight coefficient matrices, W 1∈ R d×d , W 2 ∈ R dr×d , R denotes a real matrix, d denotes a core feature dimension, dr denotes an internal feature transformation dimension of FNN .​ 4. The HRRP target recognition method based on integrated clutter suppression of the adversarial thought according to claim 3, characterized in that: The expression of the layer normalization operation in S5.7 is as follows: （12）， wherein, M n for the first n layer attention normalization operation, LN () represents LN function.

5. The HRRP target recognition method based on integrated clutter suppression of the adversarial thought according to claim 4, characterized in that: In S5.8 The operation formula of the operation is: （13）， （14）， wherein y represents a target category, and has a dimension of , s represents a residual after a normalization operation of the i-th layer attention, and has a dimension of n , , M n performing a vectorization operation on a last layer of the i-th layer encoding, n , Flatten representing flattening, and respectively represent weight matrix coefficients, and , , representing a dimension of the weight matrix coefficients, representing a dimension of the weight matrix coefficients, K representing a total number of items of the input HRRP target data.

Images