CN112114295B

CN112114295B - Method and system for identifying all-polarized radar target

Info

Publication number: CN112114295B
Application number: CN202010742966.8A
Authority: CN
Inventors: 付哲泉; 但波; 刘瑜; 康家方; 谭大宁; 张军涛; 王旭坤; 林迅
Original assignee: Naval Aeronautical University
Current assignee: Naval Aeronautical University
Priority date: 2020-07-29
Filing date: 2020-07-29
Publication date: 2023-12-22
Anticipated expiration: 2040-07-29
Also published as: CN112114295A

Abstract

The invention discloses a method and a system for identifying a full-polarization radar target. The method comprises the following steps: decomposing polarized electromagnetic waves into horizontal polarized electromagnetic waves and vertical polarized electromagnetic waves which are perpendicular to each other, determining a polarized scattering matrix according to the relation between an incident field and a scattering field of the electromagnetic waves, and performing feature extraction on the polarized scattering matrix by adopting a Pauli decomposition method, an H, alpha, A decomposition method and a feature extraction method of structural similarity parameters to obtain a first feature vector, a second feature vector and a third feature vector of target full-polarization HRRP data; and combining the first feature vector, the second feature vector and the third feature vector with the neural network to further realize the identification of the all-polarized radar target. By adopting the method and the system, the ship target HRRP information of the four polarization channels is utilized to extract the target full polarization characteristics, so that more characteristics of the target full polarization HRRP can be reserved, the accuracy of target identification is improved, and the classification effect is improved.

Description

Method and system for identifying all-polarized radar target

Technical Field

The invention relates to the technical field of radar target recognition, in particular to a method and a system for recognizing a full-polarization radar target.

Background

With the continuous development of radar technology, researchers want to obtain more effective information of targets for classification and identification of the targets, and the radar seeker of the traditional system is difficult to meet the requirement of current stage diversification. Polarization domain information in radar target echo signals is another important feature for target recognition, which follows time domain, frequency domain, etc.

Advances in polarization-related technology have enabled the use of fully polarized system radar directors to enable target identification based on polarization characteristics. Since the size of the ship target is far greater than the wavelength of the electromagnetic wave of the radar guide head, the ship target needs to be regarded as an ultra-electric large-size target when analyzing the scattering characteristics of the ship to the radar electromagnetic wave. The polarization characteristics of echo signals can be changed by vector superposition of electromagnetic waves reflected by different parts of the ship in space, so that the polarization characteristics of the ship target echo signals have direct relation with the shape, the gesture, the surface coating and the like of the ship. Polarization domain information is one of the indispensable characteristics for completely describing ship target echo signals.

Compared with the traditional monopole system radar, the full-polarization system radar has the following advantages:

(1) The radar with the full polarization system can obtain richer scattering information of the target, and relevant physical characteristics of the target including shapes, materials and the like can be estimated according to the scattering information;

(2) The full polarization information of the target echo is utilized to extract the polarization characteristics, so that the influence of a target attitude angle and the like on target identification can be reduced, and the classification capability is improved;

(3) The full-polarization system radar can utilize different receiving and transmitting polarization states to detect different types of targets more comprehensively and flexibly, and the limitation of single polarization state of electromagnetic waves of the single-polarization system radar is overcome;

(4) The full-polarization system radar can select a proper transceiving polarization mode according to the requirement, so that the anti-interference and sea clutter resistance of the radar for detecting targets are improved.

The high-resolution radar technology enables the description of the echo signals on the target information to be more accurate, and the polarization domain information enables the description of the echo signals on the target information to be more comprehensive. Radar target recognition, which combines high resolution technology and full polarization information, has received extensive attention from students and has become an important direction for the development of radar signal processing and target recognition technology.

The high resolution range profile (HRRP, high Resolution Range Profile) is the vector sum of the projection of the target scattering point sub-echoes obtained by using the broadband signal on the radar ray, and HRRP provides the distribution condition of the target scattering points along the range direction and is an important structural feature of the target. The effective separability characteristics of the HRRP are extracted, and the problems of structural similarity among different ships and amplitude, translation, attitude angle sensitivity and the like caused by high sensitivity of buildings on the ships to electromagnetic wave scattering can be solved. The current target identification research based on HRRP polarization feature extraction of the full-polarization broadband system radar mainly comprises the following two aspects:

(1) And directly combining HRRP information and polarization information of the target to construct a distance-polarization two-dimensional characteristic, and then adopting different classification and identification methods for identification.

(2) And extracting effective polarization characteristics of the target from the polarization scattering matrix based on the polarization target decomposition theory for target identification.

The research in the two aspects mainly focuses on extracting the characteristic with strong separability from the target full-polarization HRRP, and when a part of scholars calculate the coherence matrix, all the distance units contained in each element of the coherence matrix are averaged, namely all the distance units are used as measurement scales, so that the dimension of the characteristic space and the calculation complexity in the data processing process are greatly reduced, but the specific characteristic of each distance unit cannot be reserved, and the accuracy of target identification is low.

Disclosure of Invention

The invention aims to provide a full-polarization radar target identification method and system, which utilize ship target HRRP information of four polarization channels to extract target full-polarization characteristics, can reserve more characteristics of target full-polarization HRRP, improve the accuracy of target identification and improve the classification effect.

In order to achieve the above object, the present invention provides the following solutions:

A method for fully polarized radar target identification, comprising:

acquiring polarized electromagnetic waves, and decomposing the polarized electromagnetic waves into horizontal polarized electromagnetic waves and vertical polarized electromagnetic waves which are perpendicular to each other;

determining a polarization scattering matrix according to the relation between the incident field of the horizontally polarized electromagnetic wave and the scattering field of the horizontally polarized electromagnetic wave and the relation between the incident field of the vertically polarized electromagnetic wave and the scattering field of the vertically polarized electromagnetic wave; the elements of the polarization scattering matrix are respectively a horizontal co-polarization term, a vertical co-polarization term and two cross-polarization terms corresponding to HRRP data of the four polarization channels;

extracting features of the polarized scattering matrix by using a Pauli decomposition method to obtain a first feature vector of target full-polarization HRRP data; the first feature vector reflects the duty ratio of the scattering type in the target echo;

performing feature extraction on the polarized scattering matrix by adopting an H, alpha and A decomposition method to obtain a second feature vector of target full-polarization HRRP data; the second feature vector reflects the probability of the scattering mechanism occurring;

performing feature extraction on the polarized scattering matrix by adopting a feature extraction method of structural similarity parameters to obtain a third feature vector of the target full-polarization HRRP data; the third feature vector reflects the similarity of scattering features between the two targets;

And training the neural network model by taking the first feature vector, the second feature vector and the third feature vector of the target full-polarization HRRP data as input and taking the radar target class as output to obtain a trained neural network model, and carrying out full-polarization radar target identification according to the trained neural network model.

Optionally, the determining a polarization scattering matrix according to the relationship between the incident field of the horizontally polarized electromagnetic wave and the scattering field of the horizontally polarized electromagnetic wave and the relationship between the incident field of the vertically polarized electromagnetic wave and the scattering field of the vertically polarized electromagnetic wave specifically includes:

the polarization scattering matrix is determined according to the following formula:

wherein S is a polarization scattering matrix,is the incident field of horizontally polarized electromagnetic waves, +.>Is the incident field of a vertically polarized electromagnetic wave, +.>Is the scattering field of horizontally polarized electromagnetic waves, +.>Is a scattering field of vertically polarized electromagnetic wave S _HH Is a horizontal co-polarized term, S _VV Is a vertical co-polarized term S _HV For the first cross-polarized term, S _VH Is the second cross-polarized term.

Optionally, the feature extraction is performed on the polarization scattering matrix by using a Pauli decomposition method to obtain a first feature vector of the target full-polarization HRRP data, which specifically includes:

Representing the polarization scattering matrix with Pauli base to obtain S=aX ₀ +bX ₁ +cX ₂ And determining a, b, c as a first feature vector; wherein X is ₀ 、X ₁ And X ₂ The first term, the second term and the third term of Pauli base respectively, and a, b and c are the specific gravities of three scattering types in the target echo respectively;

HRRP data of four polarized channels is obtained; HRRP data of the four polarization channels are s respectively _HH (n)、s _HV (n)、s _VH (n) and s _VV (N), wherein n=0, 1,2,..n-1, N is the number of radar range units, s, of HRRP data _HH (n) is the nth radar range orderHorizontal co-polarized terms of element s _HV (n) a first cross-polarized term, s, which is the nth radar range bin _VH (n) a second cross-polarized term, s, being the nth radar range bin _VV (n) is the vertical co-polarized term of the nth radar range bin;

determining the first feature vector according to HRRP data of the four polarized channels by adopting the following formula:

wherein a (n), b (n) and c (n) are all first feature vectors of the nth radar range cell.

Optionally, the feature extraction is performed on the polarization scattering matrix by using an H, α, a decomposition method to obtain a second feature vector of the target full-polarization HRRP data, which specifically includes:

determining a polarized scattering vector of each radar distance unit according to the first characteristic vector; the polarized scatter vector k (n) for the nth radar range bin is as follows:

Determining a coherence matrix of each radar distance unit according to the polarized scattering vector; the coherence matrix T (n) of the nth radar range bin is as follows:

T(n)＝k(n)·k ^H (n)

performing eigenvalue decomposition on the coherence matrix to obtainAnd lambda is taken as _j (n) determining a j-th eigenvalue of the coherence matrix of the n-th radar range bin; wherein mu _j (n) is lambda _j (n) corresponding feature vectors;

calculating the probability of each scattering mechanism according to the characteristic values; probability P of occurrence of the jth scattering mechanism of the nth radar range bin _j (n) is as follows:

respectively calculating scattering entropy, scattering angle and anisotropy according to the probability of occurrence of a scattering mechanism, and determining the scattering entropy, the scattering angle and the anisotropy as second feature vectors; the scattering angle comprises a dominant scattering angle, and the dominant scattering angle is the scattering angle corresponding to the maximum value in the eigenvalue of the coherent matrix;

wherein,

wherein H (n) is the scattering entropy of the nth radar range bin, alpha (n) is the scattering angle of the nth radar range bin, A (n) is the anisotropy of the nth radar range bin, alpha _j (n) is a scattering angle corresponding to the jth eigenvalue of the nth radar range bin.

Optionally, the feature extraction method using structural similarity parameters performs feature extraction on the polarization scattering matrix to obtain a third feature vector of the target full-polarization HRRP data, and specifically includes:

Respectively determining a polarized scattering vector of the target and a polarized scattering vector of the standard body; the standard body comprises a flat plate, a dihedral angle, a horizontal dipole, a cylinder, a left screw body and a right screw body;

calculating similarity parameters of the target and each standard body according to the polarized scattering vector of the target and the polarized scattering vector of the standard body, and determining the similarity parameters of the target and each standard body as a third feature vector; wherein, the formula for calculating the similarity parameter is as follows:

wherein r (S) ₁ ,S ₂ ) S is the similarity parameter between polarization scattering matrices ₁ For the polarization scattering matrix of the target, S ₂ Polarization scattering matrix k being standard body ₁ For the polarized scattering vector, k of the target ₂ Is the polarized scattering vector of the standard body.

The invention also provides a full polarization radar target recognition system, which comprises:

the electromagnetic wave decomposition module is used for acquiring polarized electromagnetic waves and decomposing the polarized electromagnetic waves into horizontal polarized electromagnetic waves and vertical polarized electromagnetic waves which are perpendicular to each other;

a polarized scattering matrix determining module, configured to determine a polarized scattering matrix according to a relationship between an incident field of the horizontally polarized electromagnetic wave and a scattered field of the horizontally polarized electromagnetic wave, and according to a relationship between an incident field of the vertically polarized electromagnetic wave and a scattered field of the vertically polarized electromagnetic wave; the elements of the polarization scattering matrix are respectively a horizontal co-polarization term, a vertical co-polarization term and two cross-polarization terms corresponding to HRRP data of the four polarization channels;

The first feature vector determining module is used for extracting features of the polarized scattering matrix by using a Pauli decomposition method to obtain a first feature vector of target full-polarization HRRP data; the first feature vector reflects the duty ratio of the scattering type in the target echo;

the second feature vector determining module is used for extracting features of the polarized scattering matrix by adopting an H, alpha and A decomposition method to obtain a second feature vector of the target full-polarization HRRP data; the second feature vector reflects the probability of the scattering mechanism occurring;

the third feature vector determining module is used for carrying out feature extraction on the polarized scattering matrix by adopting a feature extraction method of the structural similarity parameter to obtain a third feature vector of the target full-polarization HRRP data; the third feature vector reflects the similarity of scattering features between the two targets;

the full-polarization radar target recognition module is used for training the neural network model by taking a first feature vector, a second feature vector and a third feature vector of the target full-polarization HRRP data as input and taking a radar target class as output to obtain a trained neural network model, and carrying out full-polarization radar target recognition according to the trained neural network model.

Optionally, the polarization scattering matrix determining module specifically includes:

a polarized scattering matrix determining unit for determining a polarized scattering matrix according to the following formula:

Optionally, the first feature vector determining module specifically includes:

a polarized scattering matrix processing unit, configured to represent the polarized scattering matrix with Pauli base, to obtain S=aX ₀ +bX ₁ +cX ₂ And determining a, b, c as a first feature vector; wherein X is ₀ 、X ₁ And X ₂ The first term, the second term and the third term of Pauli base respectively, and a, b and c are the specific gravities of three scattering types in the target echo respectively;

the HRRP data acquisition unit is used for acquiring the HRRP data of the four polarized channels; HRRP data of the four polarization channels are s respectively _HH (n)、s _HV (n)、s _VH (n) and s _VV (N), wherein n=0, 1,2,..n-1, N is the number of radar range units, s, of HRRP data _HH (n) is the horizontal co-polarization term of the nth radar range bin, s _HV (n) a first cross-polarized term, s, which is the nth radar range bin _VH (n) a second cross-polarized term, s, being the nth radar range bin _VV (n) is the vertical co-polarized term of the nth radar range bin;

a first feature vector determining unit, configured to determine the first feature vector according to HRRP data of the four polarized channels by using the following formula:

Optionally, the second feature vector determining module specifically includes:

a polarized scattering vector determining unit, configured to determine a polarized scattering vector of each radar range cell according to the first feature vector; the polarized scatter vector k (n) for the nth radar range bin is as follows:

a coherence matrix determining unit, configured to determine a coherence matrix of each radar distance unit according to the polarized scattering vector; the coherence matrix T (n) of the nth radar range bin is as follows:

T(n)＝k(n)·k ^H (n)

a eigenvalue decomposition unit for decomposing eigenvalues of the coherence matrix to obtainAnd lambda is taken as _j (n) determining a j-th eigenvalue of the coherence matrix of the n-th radar range bin; wherein mu _j (n) is lambda _j (n) corresponding feature vectors;

a probability calculation unit for calculating the probability of each scattering mechanism according to the characteristic value; probability P of occurrence of the jth scattering mechanism of the nth radar range bin _j (n) is as follows:

a second feature vector determining unit configured to calculate a scattering entropy, a scattering angle, and an anisotropy degree, respectively, according to a probability that a scattering mechanism occurs, and determine the scattering entropy, the scattering angle, and the anisotropy degree as second feature vectors; the scattering angle comprises a dominant scattering angle, and the dominant scattering angle is the scattering angle corresponding to the maximum value in the eigenvalue of the coherent matrix;

wherein,

Optionally, the third feature vector determining module specifically includes:

a polarized scattering vector determination unit for determining a polarized scattering vector of the target and a polarized scattering vector of the standard body, respectively; the standard body comprises a flat plate, a dihedral angle, a horizontal dipole, a cylinder, a left screw body and a right screw body;

A third feature vector determining unit, configured to calculate similarity parameters of the target and each standard body according to the polarized scattering vector of the target and the polarized scattering vector of the standard body, and determine the similarity parameters of the target and each standard body as a third feature vector; wherein, the formula for calculating the similarity parameter is as follows:

Compared with the prior art, the invention has the beneficial effects that:

the invention provides a full-polarization radar target identification method and system, which are characterized in that polarized electromagnetic waves are decomposed into horizontal polarized electromagnetic waves and vertical polarized electromagnetic waves which are perpendicular to each other, a polarization scattering matrix is determined according to the relation between an incident field and a scattering field of the electromagnetic waves, and the polarization scattering matrix is subjected to feature extraction by a Pauli decomposition method, an H, alpha, A decomposition method and a feature extraction method of structural similarity parameters to obtain a first feature vector, a second feature vector and a third feature vector of target full-polarization HRRP data; the first feature vector, the second feature vector and the third feature vector are all polarization features of the target, and the all polarization features of the target are combined with the neural network so as to realize the identification of the all-polarization radar target. According to the invention, the ship target HRRP information of the four polarization channels is utilized to extract the target full polarization characteristics, and meanwhile, a single distance unit is selected as a measurement scale, so that more characteristics of the target full polarization HRRP can be reserved, the target identification accuracy is improved, and the classification effect is improved.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the drawings that are needed in the embodiments will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present invention, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a flowchart of a method for identifying a fully polarized radar target according to an embodiment of the present invention;

FIG. 2 is a diagram of a full polarization radar target recognition system according to an embodiment of the present invention;

FIG. 3 is a diagram of a residual error architecture in a second embodiment of the present invention;

FIG. 4 is a diagram of a convolution module according to a second embodiment of the present disclosure;

FIG. 5 is a block diagram of a neural network model in a second embodiment of the present invention;

fig. 6 is a flowchart of object recognition based on polarization feature extraction in the second embodiment of the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

In order that the above-recited objects, features and advantages of the present invention will become more readily apparent, a more particular description of the invention will be rendered by reference to the appended drawings and appended detailed description.

Example 1

Fig. 1 is a flowchart of a method for identifying a fully polarized radar target in an embodiment of the present invention, as shown in fig. 1, a method for identifying a fully polarized radar target includes:

step 101: polarized electromagnetic waves are acquired and decomposed into horizontal polarized electromagnetic waves and vertical polarized electromagnetic waves that are perpendicular to each other.

Step 102: determining a polarization scattering matrix according to the relation between the incident field of the horizontally polarized electromagnetic wave and the scattering field of the horizontally polarized electromagnetic wave and the relation between the incident field of the vertically polarized electromagnetic wave and the scattering field of the vertically polarized electromagnetic wave; the elements of the polarized scattering matrix are respectively a horizontal co-polarized term, a vertical co-polarized term and two cross-polarized terms corresponding to HRRP data of the four polarized channels.

Step 102, specifically includes:

wherein S is a polarization scattering matrix,for horizontally polarized electromagneticWave incident field->Is the incident field of a vertically polarized electromagnetic wave, +.>Is the scattering field of horizontally polarized electromagnetic waves, +.>Is a scattering field of vertically polarized electromagnetic wave S _HH Is a horizontal co-polarized term, S _VV Is a vertical co-polarized term S _HV For the first cross-polarized term, S _VH Is the second cross-polarized term.

Step 103: extracting features of the polarized scattering matrix by using Pauli decomposition method to obtain a first feature vector of target full-polarization HRRP data; the first feature vector reflects the duty cycle of the scattering type in the target echo.

Step 103, specifically includes:

representation of the polarized scattering matrix using Pauli base, resulting in S=aX ₀ +bX ₁ +cX ₂ And determining a, b, c as a first feature vector; wherein X is ₀ 、X ₁ And X ₂ The first term, the second term and the third term of Pauli base respectively, and a, b and c are the specific gravities of three scattering types in the target echo respectively;

HRRP data of four polarized channels is obtained; HRRP data for four polarized channels are s respectively _HH (n)、s _HV (n)、s _VH (n) and s _VV (N), wherein n=0, 1,2,..n-1, N is the number of radar range units, s, of HRRP data _HH (n) is the horizontal co-polarization term of the nth radar range bin, s _HV (n) a first cross-polarized term, s, which is the nth radar range bin _VH (n) a second cross-polarized term, s, being the nth radar range bin _VV (n) is the vertical co-polarized term of the nth radar range bin;

the first feature vector is determined from HRRP data for the four polarized channels using the following formula:

Step 104: performing feature extraction on the polarization scattering matrix by adopting an H, alpha and A decomposition method to obtain a second feature vector of the target full-polarization HRRP data; the second feature vector reflects the probability of the scattering mechanism occurring.

Step 104 specifically includes:

T(n)＝k(n)·k ^H (n)

eigenvalue decomposition is carried out on the coherence matrix to obtainAnd lambda is taken as _j (n) determining a j-th eigenvalue of the coherence matrix of the n-th radar range bin; wherein mu _j (n) is lambda _j (n) corresponding feature vectors;

calculating the probability of each scattering mechanism according to the characteristic value; probability P of occurrence of the jth scattering mechanism of the nth radar range bin _j (n) is as follows:

wherein,

Step 105: performing feature extraction on the polarized scattering matrix by adopting a feature extraction method of the structural similarity parameters to obtain a third feature vector of the target full-polarization HRRP data; the third feature vector reflects the similarity of scattering features between the two objects.

Step 105 specifically includes:

Step 106: and training the neural network model by taking the first feature vector, the second feature vector and the third feature vector of the target full-polarization HRRP data as input and the radar target class as output to obtain a trained neural network model, and carrying out full-polarization radar target identification according to the trained neural network model.

Fig. 2 is a block diagram of a full polarization radar target recognition system according to an embodiment of the present invention. As shown in fig. 2, a full polarization radar target recognition system includes:

an electromagnetic wave decomposition module 201 for acquiring polarized electromagnetic waves and decomposing the polarized electromagnetic waves into horizontal polarized electromagnetic waves and vertical polarized electromagnetic waves perpendicular to each other.

A polarized scattering matrix determining module 202, configured to determine a polarized scattering matrix according to a relationship between an incident field of the horizontally polarized electromagnetic wave and a scattered field of the horizontally polarized electromagnetic wave, and according to a relationship between an incident field of the vertically polarized electromagnetic wave and a scattered field of the vertically polarized electromagnetic wave; the elements of the polarized scattering matrix are respectively a horizontal co-polarized term, a vertical co-polarized term and two cross-polarized terms corresponding to HRRP data of the four polarized channels.

The polarization scattering matrix determining module 202 specifically includes:

The first feature vector determining module 203 is configured to perform feature extraction on the polarization scattering matrix by using a Pauli decomposition method, so as to obtain a first feature vector of target full-polarization HRRP data; the first feature vector reflects the duty cycle of the scattering type in the target echo.

The first feature vector determining module 203 specifically includes:

a polarized scattering matrix processing unit, configured to represent the polarized scattering matrix with Pauli base to obtain S=aX ₀ +bX ₁ +cX ₂ And determining a, b, c as a first feature vector; wherein X is ₀ 、X ₁ And X ₂ The first term, the second term and the third term of Pauli base respectively, and a, b and c are the specific gravities of three scattering types in the target echo respectively;

the HRRP data acquisition unit is used for acquiring the HRRP data of the four polarized channels; HRRP data for four polarized channels are s respectively _HH (n)、s _HV (n)、s _VH (n) and s _VV (N), wherein n=0, 1,2,..n-1, N is the number of radar range units, s, of HRRP data _HH (n) water as the nth radar range binFlattened term, s _HV (n) a first cross-polarized term, s, which is the nth radar range bin _VH (n) a second cross-polarized term, s, being the nth radar range bin _VV (n) is the vertical co-polarized term of the nth radar range bin;

a first feature vector determining unit, configured to determine a first feature vector according to HRRP data of four polarized channels using the following formula:

/>

The second feature vector determining module 204 is configured to perform feature extraction on the polarized scattering matrix by using an H, α, a decomposition method, so as to obtain a second feature vector of the target full-polarization HRRP data; the second feature vector reflects the probability of the scattering mechanism occurring.

The second feature vector determining module 204 specifically includes:

a polarized scattering vector determining unit for determining a polarized scattering vector of each radar range cell according to the first feature vector; the polarized scatter vector k (n) for the nth radar range bin is as follows:

the coherent matrix determining unit is used for determining a coherent matrix of each radar distance unit according to the polarized scattering vector; the coherence matrix T (n) of the nth radar range bin is as follows:

T(n)＝k(n)·k ^H (n)

a probability calculation unit for calculating the probability of occurrence of each scattering mechanism according to the characteristic value; probability P of occurrence of the jth scattering mechanism of the nth radar range bin _j (n) is as follows:

a second feature vector determination unit configured to calculate a scattering entropy, a scattering angle, and an anisotropy degree, respectively, according to a probability that a scattering mechanism occurs, and determine the scattering entropy, the scattering angle, and the anisotropy degree as second feature vectors; the scattering angle comprises a dominant scattering angle, and the dominant scattering angle is the scattering angle corresponding to the maximum value in the eigenvalue of the coherent matrix;

Wherein,

The third feature vector determining module 205 is configured to perform feature extraction on the polarized scattering matrix by using a feature extraction method of the structural similarity parameter, so as to obtain a third feature vector of the target full-polarization HRRP data; the third feature vector reflects the similarity of scattering features between the two objects.

The third feature vector determining module specifically includes:

The full-polarization radar target recognition module 206 is configured to train the neural network model with the first feature vector, the second feature vector, and the third feature vector of the target full-polarization HRRP data as input and the radar target class as output, obtain a trained neural network model, and perform full-polarization radar target recognition according to the trained neural network model.

Example two

Polarization of an electromagnetic wave is defined as the way in which the orientation of the electromagnetic wave's electric field vector E in space changes over time. For planar electromagnetic waves, an electromagnetic wave e of arbitrary polarization _HV Can be described as being decomposed into two mutually perpendicular linear polarization components, namely:

in the formula (1), the lower subscript is H andv represents a horizontal polarization component and a vertical polarization component; a, a _H Andrepresenting the amplitude and phase, respectively, in the horizontally polarized component of the planar electromagnetic wave; a, a _V And->The amplitude and phase in the vertically polarized component of the planar electromagnetic wave are represented, respectively.

The polarization scattering matrix S represents the incident field E of the electromagnetic wave at the target ⁱ And scattered field E ^s The relationship between them can be expressed as:

E ^s ＝SE ⁱ (2)

when the radar seeker detects a target, the distance between the radar and the target is far, so that the electromagnetic wave at the target can be regarded as a plane wave. Under the horizontal polarization and vertical polarization basis, the polarization scattering matrix of the target can be represented by a 2×2 matrix, and then the expression (2) becomes:

as can be seen from the formula (3), when the radar seeker of the broadband full-polarization system detects the target, HRRP information under 4 receiving and transmitting polarization states, namely HH, HV, VH, VV, can be obtained. In the polarization scattering matrix S, S _HH And S is _VV Is homopolar item, S _HV And S is _VH Is a cross-polarized term. For the target of the linear scatterer, it can be known from the reciprocity theorem that its single-station scattering matrix is symmetrical, i.e. S _HV ＝S _VH . (compared with the HRRP obtained by the unipolar system radar, the full-polarization HRRP can obtain richer scattering information of the targets, and the scattering characteristic difference between different targets is furthest reflected in a vector form。)

Feature extraction is performed by a polarization scattering matrix S, including Pauli decomposition, H, alpha, A, alpha ₁ The physical significance of the extracted features is clear in 3 aspects of decomposition and based on structural similarity parameters, and the target characteristics can be reflected from different angles.

1) Pauli decomposition-based polar scattering matrix feature extraction

Polar scattering matrix feature extraction based on Pauli decomposition is a method of coherent decomposition, where the total backscatter of the target is represented by a weighted summation of several typical scattering mechanisms. Other common coherent decomposition methods are also Krogager decomposition-based methods and Cameron decomposition-based methods, which are more intuitive than other methods.

Under a linear orthogonal basis for vertical and horizontal polarization, the Pauli group can be expressed as:

in the case where the reciprocity theorem is satisfied (S _HV ＝S _VH ) Pauli groups can now be reduced to X ₀ 、X ₁ 、X ₂ Three terms, namely, the polarization scattering matrix, are expressed as:

the combination of formula (4) and formula (5) can give:

a. b and c are complex numbers, and respectively represent the proportion of three scattering types in the target echo, |a| ² 、|b| ² And |c| ² The three scattering types of energy contained in the target echo are represented respectively.

Pauli groups are a complete set of orthogonal groups and have the advantage of simple form. Pauli decomposition can also be performed on HRRP data with noise, namely the Pauli decomposition has certain anti-interference capability. It should also be noted that Pauli decomposition only corresponds to both odd and even scattering types, and is not able to describe more complex scattering characteristics of the target. Let the HRRP data of each polarized channel be s _HH (n)、s _HV (n)、s _VH (n)、s _VV (N), where n=0, 1,2,..n-1, N is the number of distance units of HRRP data. When a single distance cell is selected as a metric, polarized scattering features a, b, c based on Pauli decomposition can be derived by equation (9).

Through the data processing, 3 feature vectors of a, b and c of the target full-polarization HRRP data can be obtained.

2) Based on H, alpha, A, alpha ₁ Decomposed polarization scattering matrix feature extraction

Representing the a, b, c characteristics as a vector, and coefficientThe polarization scattering vector can be obtained by extracting the component vectors of the corresponding positions of the rest contents:

while the coherence matrix can be defined as follows:

T＝k·k ^H (11)

in the formula (11), T is a semi-positive Hermit matrix, and can be obtained by performing characteristic decomposition on T:

in the formula (12), T _i ＝λ _i μ _i ·μ _i ^H (i=1, 2, 3) corresponds to one steady state target, respectively; lambda (lambda) _i Is the ith eigenvalue of the coherence matrix T, and lambda ₁ ≥λ ₂ ≥λ ₃ ≥0；μ _i As a characteristic value lambda _i The corresponding feature vectors are expressed as:

in the formula (13), alpha _i And beta _i Respectively representing a scattering mechanism and an orientation angle of the target;represent S _HH +S _VV Is a phase of (2); delta _i Represent S _HH +S _VV And S is equal to _HH -S _VV A phase difference between them; gamma ray _i Then represent S _HH +S _VV And S is equal to _HV Phase difference between them.

According to probability of occurrence of each scattering mechanismThe following 3 characteristic parameters can be defined:

H is scattering entropy, which describes the proportion of different scattering mechanisms in the total scattering process, and also describes the randomness of target scattering, and the value interval is [0,1]. When the value of H increases from 0 to 1, the target scattering changes from a fully polarized state to a fully unpolarized state. Alpha is a scattering angle, which describes the degree of freedom inside the target and reflects the main scattering mechanism of the target, and the value interval is [0 DEG, 90 DEG ]. When the value of alpha increases from 0 deg. to 90 deg., the main scattering mechanism of the target changes from isotropic surface scattering to dihedral scattering. A is the degree of anisotropy, which describes the degree of anisotropy of the scattering of the target.

α ₁ Is the scattering angle corresponding to the maximum value in the eigenvalue of the coherence matrix, called dominant scattering angle, which represents the dominant scattering mechanism of the target echo signal. And thus may be retained as an important feature.

H, α, a, α for wideband full-polarization system radar target HRRP data when a single range bin is selected as a metric ₁ The feature extraction process can be summarized as shown in table 1:

table 1H, α, A, α ₁ Feature extraction process

The H, alpha, A and alpha of the target full-polarization HRRP data can be obtained through the data processing ₁ A total of 4 feature vectors.

3) Polarization scattering matrix feature extraction based on structural similarity parameters

The feature extraction of the polarized scattering matrix based on the structural similarity parameter reflects the similarity of scattering features between two targets, and the parameter is irrelevant to the attitude angle of the targets and the power of echo signals.

From the above, the polarized scattering vector is obtainedThe similarity parameter between different scattering matrixes can be obtained according to the polarized scattering vector, and the similarity parameter r is defined as follows:

in equation (17), vector k ₁ And k ₂ Respectively polarized scattering matrix S ₁ And S is ₂ Is used for the polarization scattering vector of the (c).Representing the sum of squares of the individual elemental modes of the polarized scatter vector.

Combining formula (17) with the polarization scattering matrix of the common standard body can obtain the similarity parameter of the target and the standard body. The polarization scattering matrix of a common standard body and the polarization scattering vector under Pauli are shown in Table 2.

TABLE 2 polarization scattering matrix of common Standard volume and polarization scattering vector under Pauli

Bringing the polarized scattering vector of each standard body under Pauli into formula (17) to obtain the structural similarity parameters of the target and different standard bodies, wherein the structural similarity parameters are respectively as follows:

(1) Structural similarity parameters of target and plate:

(2) Structural similarity parameters of target and dihedral angle:

(3) Structural similarity parameters of target and horizontal dipole:

(4) Structural similarity parameters of target and cylinder:

(5) Structural similarity parameters of target and left screw:

(6) Structural similarity parameters of target and right screw:

as before, a single distance unit is selected as a metric, the structural similarity parameters of the targets are sequentially calculated for the polarized scattering matrix of each distance unit, and the HRRP data s of each polarized channel is calculated _HH (n)、s _HV (n)、s _VH (n)、s _VV (n) structural similarity parameters of the target and the plate are:

the structural similarity parameters of the target and other standard bodies can be obtained by referring to formulas (19 to 23). Through the above dataR of target full polarization HRRP data can be obtained through processing _plane 、r _dihedral 、r _dipole 、r _cylinder 、r _lhelix 、r _rhelix A total of 6 feature vectors.

By performing the feature extraction of the above 3 modes on ship target full polarization HRRP data, each ship target can finally obtain 13 feature vectors, as shown in table 3.

TABLE 3 set of target feature vectors for ships

Table 1Ship targetfeature vector set

/>

By Pauli decomposition of the polarized scattering matrix, H, alpha, A, alpha ₁ The feature extraction of 3 aspects of decomposition and structure similarity parameters can obtain the feature vectors with clear physical meaning, and the target characteristics can be reflected from different angles.

The depth of the neural network is critical, and the deep convolutional neural network can extract and fuse features of different layers for end-to-end target recognition. However, the deepening of the network layer number brings about the problem of saturation of the recognition accuracy, and the residual structure is generally introduced to overcome the problem, and the structure is shown in fig. 3.

The residual blocks in the residual structure consist of convolution layers, the number of which is 2 in the figure. The output of the residual structure is the input feature added to the output of the last convolutional layer, represented by equation (25)

x _l+1 ＝F(x _l )+x _l (25)

Wherein x is _l 、x _l+1 Respectively representing input and output characteristics of a first layer residual structure; f (x) _l ) Representing a mapping of the residual block.

Studies have shown that mapping F (x _l ) Replace the required mapping F #x _l )+x _l The problem of saturation of the deep network identification accuracy can be effectively solved. Because in extreme cases, if the network has already extracted the best features required for classification, the residual structure can ensure the highest recognition accuracy only by performing the identity mapping of the jump connection. And for neural networks, zeroing out the residual block is more efficient than fitting an identity mapping using a multi-layer neural network.

In order to facilitate understanding of the neural network classifier used in the present invention, a specific structure of the fusion convolution module and the improved residual structure neural network proposed for ship target single polarization channel identification is briefly described herein, as shown in fig. 4. In fig. 4, mx1×n represents one-dimensional data having an input characteristic of mx1, the number of characteristic layers is N, and s is a shift step of the convolution kernel. The convolution module is set to be a highly modularized network structure, and the expandability is strong. The features extracted by the upper layer network serve as inputs to the layer network, which will go through 2 branches. In the left branch, features among different layers are first fused by using a 1×1 convolution kernel, the fused features are equally divided into a plurality of branches on the layer number, each branch has 3-layer features, each branch respectively uses a 3×1 convolution kernel to perform feature extraction, the step length is 2, the number of the feature layers is unchanged, and the dimension is halved. And then splicing the plurality of branch characteristics, determining the size of x according to the complexity of a classification task (the structure is similar to an acceptance structure, but the size and the number of convolution kernels of each branch in the acceptance are customized step by step, uniformly selecting 3X 1 small-scale convolution kernels to reduce the structural design difficulty, and guaranteeing the recognition effect.) the spliced characteristics are subjected to characteristic fusion again by using 1X 1 convolution kernels, the number of characteristic layers is increased, and the characteristics are divided into two parts according to layers so as to prepare for the characteristic fusion of two subsequent branches. The right branch directly uses the convolution check input of 1 multiplied by 1 to perform feature fusion and increase the number of feature layers, and simultaneously the features are divided into two parts according to the number of layers, and the addition and the splicing operation are performed on the features corresponding to the left branch.

The output of the convolution module is halved in characteristic dimension and doubled in layer number compared with the input. The effect of the right branch is similar to that of a residual network, so that each layer of the network module can acquire information from a loss function and an original input signal, characteristics and gradients are transferred more effectively, the utilization rate of shallow characteristics is improved, and the problems that the gradients possibly generated with the continuous deepening of the network disappear and the recognition rate is saturated are solved.

The loss function is used to measure the difference between the predicted value and the actual value, and is generally denoted by L (y_, y), where y_ represents the predicted value and y represents the actual value. ( For multi-class convolutional neural networks, softmax Loss (SL) is typically used as the Loss function. However, from a clustering perspective, SL extracted features may have an intra-class distance greater than an inter-class distance. Meanwhile, the SL extracted features are divergent in visualization, and the features are overlapped when the target categories are too many, so that the target classification is not facilitated. Many solutions to this problem have been proposed, focusing mainly on increasing the inter-class distance and decreasing the intra-class distance. )

For the functional form of Softmax, features can be urged closer together by increasing the inter-class distance by enhancing boundary constraints between different target classes. The Softmax loss function is calculated as follows:

Where x is the input of the last fully connected layer,represents the ith deep feature and belongs to the yi th category, d represents the dimension of the feature; />Weight matrix representing the last fully connected layer +.>Is the j-th column of (2); />Representing the weighted result of the i-th sample target feature; m represents each batch in the network training processNumber of samples; n represents the number of categories of the target.

The loss function fuses boundary constraint and center clustering, and the specific expression is as follows:

wherein s is used to control the cosine distance between features, which represents the similarity of features; μ is used to control the magnitude of the distance between feature edges;represents the y _i Feature center of individual target class, +.>And continuously updated as each batch of data depth features changes. L (L) _AMS And the distance between classes is constrained by introducing psi (theta) =cos theta-mu, and the recognition effect is improved by increasing the distance between the classes of the features. L (L) _C A class center is constructed for each feature of the target class, and target features far from the class center are penalized. The intra-class distance of the target features is more compact, and the effects of reducing the intra-class distance and increasing the inter-class distance are achieved. The update formula for the category center is as follows:

when y is _i For the j-th class of targets, the identification is correct, when delta (·) is equal to 1, otherwise delta (·) is equal to 0. At the joint loss function L _AMSC Under constraints, the training and learning process of the model can be summarized as shown in table 4:

table 4 model training and learning process

The model block diagram of the invention is shown in fig. 5, and mainly comprises an initial convolution layer, a plurality of convolution modules with the same topological structure which are connected in sequence and two final full connection layers. The dimension of the latter full-connection layer is 2, so that the characteristics extracted by the model can be visualized, and the clustering effect of the characteristics can be analyzed.

The numbers in brackets indicate the data dimensions after the HRRP sample passes through the layer, as in fig. 4. The output data dimension of the plurality of convolution modules and the first fully-connected layer is determined according to the number of the convolution modules. The result of the last output layer is one-dimensional data corresponding to the target class, here corresponding to the target class number 3. The initial convolution layer in the model selects a one-dimensional convolution kernel with the scale of 7 multiplied by 1, and the first layer of the network selects a convolution kernel with a relatively large scale, so that the corresponding features such as contours, textures and the like in the target HRRP data can be extracted. And carrying out batch normalization and Relu activation operation on the extracted features after each convolution operation in the model.

The model parameters remain unchanged when the ship target is subjected to unipolar HRRP data, but because the invention also processes ship target full-polarization HRRP data, after feature extraction is carried out on the full-polarization HRRP data, each data set HRRP is changed from one-dimensional data to two-dimensional data. Therefore, when processing ship target full-polarization HRRP data, the structure of the neural network model needs to be correspondingly adjusted. The convolution kernels in the network structure are all expanded into two-dimensional convolution kernels, the convolution kernel size of the initial convolution layer is changed from 7×1 to 7×3, and the convolution kernel size in the convolution module is changed from 3×1 to 3×3. The step size of the convolution kernel is changed from one dimension to two dimensions correspondingly, and the step size of the convolution kernel is set to s= (2, 1), namely the step size of the convolution kernel moving in the direction of the distance unit is 2, and the step size of the convolution kernel moving between data channels is 1. The convolution kernel corresponding step size of 1×1 in the convolution module is set to s= (1, 1). The pooling blocks in the pooling layer also correspond to a change from one dimension 3×1 to two dimension 3×3, with a step size of s= (2, 1). In summary, a flowchart of the method of the present invention is shown in fig. 6.

Aiming at ship target full-polarization HRRP data, the invention provides a high-efficiency expandable neural network by improving a residual error structure, wherein the structure is further expanded on the basis of unipolar HRRP data processing, a convolution kernel in the network structure is expanded from a one-dimensional convolution kernel to a two-dimensional convolution kernel, and the convolution kernel size of an initial convolution layer and the convolution kernel size in a convolution module are correspondingly expanded. The method can solve the problems that gradient explosion and disappearance and network identification rate saturation can occur in the back propagation process of the convolutional neural network during training along with the increase of the network depth. Meanwhile, a novel joint loss function is provided by fusing the central cluster and the feature boundary constraint, so that the recognition accuracy is further improved. Through the design of a modularized structure, the model can be efficiently expanded to adapt to classification tasks with different difficulties when aiming at single-polarization or full-polarization target HRRP data. Aiming at the problem that when the separable characteristics are extracted from the target full-polarization HRRP, the specific characteristics of each distance unit cannot be reserved by using all the distance units as measurement scales. According to the method, a single distance unit is selected as a measurement scale when ship target HRRP information of four polarization channels is comprehensively utilized, and specific characteristics of each distance unit can be reserved to the maximum extent.

The principles and embodiments of the present invention have been described herein with reference to specific examples, the description of which is intended only to assist in understanding the methods of the present invention and the core ideas thereof; also, it is within the scope of the present invention to be modified by those of ordinary skill in the art in light of the present teachings. In summary, the present description should not be construed as limiting the invention.

Claims

1. A method for fully polarized radar target identification, comprising:

extracting features of the polarized scattering matrix by using a Pauli decomposition method to obtain a first feature vector of target full-polarization HRRP data; the first feature vector reflects the duty ratio of the scattering type in the target echo; assuming HRRP data of each polarization channel is sHH (N), sHV (N), sHH (N), sVV (N), where n=0, 1,2,..n-1, N is the number of distance units of HRRP data; selecting a single distance cell as a metric;

Performing feature extraction on the polarized scattering matrix by adopting an H, alpha and A decomposition method to obtain a second feature vector of target full-polarization HRRP data; the second feature vector reflects the probability of the scattering mechanism occurring; selecting a single distance cell as a metric;

training a neural network model by taking a first feature vector, a second feature vector and a third feature vector of the target full-polarization HRRP data as input and taking a radar target class as output to obtain a trained neural network model, and carrying out full-polarization radar target identification according to the trained neural network model, wherein convolution kernels in a network structure of the neural network model are all expanded into two-dimensional convolution kernels, the convolution kernel size of an initial convolution layer is changed from 7×1 to 7×3, the convolution kernel size in a convolution module is changed from 3×1 to 3×3, the step size of the convolution kernels is correspondingly changed from one dimension to two dimension, the step size of the convolution kernels is set to s= (2, 1), namely the step size of the convolution kernels moving in the direction of a distance unit is set to 2, the step size of the convolution kernels moving among data channels is set to 1, the convolution kernels with the size of 1×1 in the convolution module is correspondingly changed from one dimension 3×1 to two dimension 3×3, and the step size of the pooling block in the pooling layer is correspondingly changed from one dimension 3×1 to two dimension 3×3, and the step size of the convolution kernels is changed to s= (2, 1); and fusing the central cluster and the feature boundary constraint, and providing a new joint loss function.

2. The method for identifying a fully polarized radar target according to claim 1, wherein the determining a polarized scattering matrix according to a relationship between the incident field of the horizontally polarized electromagnetic wave and the scattered field of the horizontally polarized electromagnetic wave and a relationship between the incident field of the vertically polarized electromagnetic wave and the scattered field of the vertically polarized electromagnetic wave specifically includes:

3. The method for identifying a fully polarized radar target according to claim 2, wherein the feature extraction is performed on the polarized scattering matrix by using a Pauli decomposition method to obtain a first feature vector of target fully polarized HRRP data, and specifically comprises:

the method for extracting the characteristics of the polarized scattering matrix by using Pauli decomposition method to obtain a first characteristic vector of target full-polarization HRRP data specifically comprises the following steps:

HRRP data of four polarized channels is obtained; HRRP data of the four polarization channels are s respectively _HH (n)、s _HV (n)、s _VH (n) and s _VV (N), wherein n=0, 1,2,..n-1, N is the number of radar range units, s, of HRRP data _HH (n) is the horizontal co-polarization term of the nth radar range bin, s _HV (n) a first cross-polarized term, s, which is the nth radar range bin _VH (n) a second cross-polarized term, s, being the nth radar range bin _VV (n) is the vertical co-polarized term of the nth radar range bin;

4. The method for identifying a fully polarized radar target according to claim 3, wherein the feature extraction is performed on the polarized scattering matrix by using an H, α, a decomposition method to obtain a second feature vector of target fully polarized HRRP data, and specifically comprises:

T(n)＝k(n)·k ^H (n)

wherein,

5. The method for identifying a fully polarized radar target according to claim 4, wherein the feature extraction method for extracting features from the polarized scattering matrix by using a structural similarity parameter, to obtain a third feature vector of target fully polarized HRRP data, specifically includes:

6. A fully polarized radar target recognition system, comprising:

the first feature vector determining module is used for extracting features of the polarized scattering matrix by using a Pauli decomposition method to obtain a first feature vector of target full-polarization HRRP data; the first feature vector reflects the duty ratio of the scattering type in the target echo; assuming HRRP data of each polarization channel is sHH (N), sHV (N), sHH (N), sVV (N), where n=0, 1,2,..n-1, N is the number of distance units of HRRP data; selecting a single distance cell as a metric;

the second feature vector determining module is used for extracting features of the polarized scattering matrix by adopting an H, alpha and A decomposition method to obtain a second feature vector of the target full-polarization HRRP data; the second feature vector reflects the probability of the scattering mechanism occurring; selecting a single distance cell as a metric;

the full-polarization radar target recognition module is used for training a neural network model by taking a first feature vector, a second feature vector and a third feature vector of target full-polarization HRRP data as input and taking a radar target class as output to obtain a trained neural network model, and carrying out full-polarization radar target recognition according to the trained neural network model, wherein convolution kernels in a network structure of the neural network model are expanded into two-dimensional convolution kernels by one-dimensional convolution kernels, the convolution kernel size of an initial convolution layer is changed from 7×1 to 7×3, the convolution kernel size in the convolution module is changed from 3×1 to 3×3, the step size of the convolution kernels is correspondingly changed from one-dimensional to two-dimensional, the step size of the convolution kernels is set to s= (2, 1), namely the convolution kernels move in the direction of a distance unit by step size of 2, the step size of the convolution kernels moving among data channels is 1, the convolution kernels in the convolution module corresponds to the step size of 1×1 is set to s= (1, 1), and the pooling blocks in the pooling layer correspondingly changed from one-dimensional 3×1 to two-dimensional 3×3, and the step size of the convolution kernels is changed from one-dimensional to s= (2, 1); and fusing the central cluster and the feature boundary constraint, and providing a new joint loss function.

7. The full polarization radar target identification system of claim 6, wherein the polarization scattering matrix determination module specifically comprises:

8. The full polarization radar target recognition system of claim 7, wherein the

The first feature vector determining module specifically includes: a polarized scattering matrix processing unit, configured to represent the polarized scattering matrix with Pauli base, to obtain S=aX ₀ +bX ₁ +cX ₂ And determining a, b, c as a first feature vector; wherein X is ₀ 、X ₁ And X ₂ The first term, the second term and the third term of Pauli base respectively, and a, b and c are the specific gravities of three scattering types in the target echo respectively;

9. The full polarization radar target identification system of claim 8, wherein the second feature vector determination module specifically comprises:

T(n)＝k(n)·k ^H (n)

wherein,

10. The full polarization radar target identification system of claim 9, wherein the third feature vector determination module specifically comprises: a polarized scattering vector determination unit for determining a polarized scattering vector of the target and a polarized scattering vector of the standard body, respectively; the standard body comprises a flat plate, a dihedral angle, a horizontal dipole, a cylinder, a left screw body and a right screw body;