A multi-feature fusion radar target recognition system and method based on ensemble learning

Abstract

The application discloses a multi-feature fusion radar target recognition system and method based on ensemble learning, and belongs to the technical field of radar signal classification. The system comprises: an angle invariant feature extraction network, which is used for inputting HRRP radar data of different angles, and fusing all single features of different angles by using a multi-layer sparse autoencoder to obtain angle invariant features of different angles; the multi-layer sparse autoencoder comprises an encoder layer, a deep encoding layer, a decoder layer and an output layer, and increases a regularization constraint on the basis of a mean square error function; an HRRP radar recognition network, which is used for classifying and recognizing the angle invariant features of different angles to form a classification result corresponding to the angle invariant features of each angle; and an integrated decision network, which is used for generating an identification result label of the HRRP radar data by means of a voting decision. The method disclosed by the application can directly recognize and classify HRRP data of different targets at different angles.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar signal classification, and particularly relates to a multi-feature fusion radar target recognition system and method based on ensemble learning. BACKGROUND

[0002] High Resolution Range Profile (HRRP) is the vector sum of target scattering point subechoes projected on the radar ray obtained by wideband radar signals, reflects the distribution of the Radar Cross Section (RCS) of a scattering body on a target along the Radar Line of Sight (RLOS) under a specific radar viewing angle, and contains important information such as the shape, orientation and distance of the target. HRRP is one-dimensional data, and compared with Synthetic Aperture Radar (SAR) and Inverse Synthetic Aperture Radar (ISAR) images, 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 existing radar target recognition method based on HRRP has recognition mismatch and failure problems when applied to spaceborne radar sea surface target recognition. The difficulties in recognizing spaceborne radar sea surface targets based on HRRP mainly lie in the following three aspects: (1) The traditional HRRP radar target recognition method excessively relies on angle features and has strong angle sensitivity. When the radar observation angle changes, the HRRP data of the target will change significantly, which greatly affects the accuracy and stability of recognition. (2) There are differences between different individuals of the same type of target. The traditional single-in-library target recognition method cannot effectively distinguish these subtle differences, often leading to recognition failure. Taking the recognition of different types of destroyers as an example, although they belong to the same category, there are subtle differences in shape, structure, etc. between individuals due to production batches, modifications, etc. The traditional method can hardly accurately distinguish these individuals according to limited features. (3) The traditional single-in-library target recognition method is limited to specific data sets and features, and has poor generalization ability. When facing new target samples or complex and variable environments, its recognition performance will be greatly reduced. SUMMARY

[0004] The application aims to provide a multi-feature fusion radar target recognition system and method based on ensemble learning, to solve the problems of strong angle sensitivity, poor generalization ability, etc. of the existing HRRP radar target recognition method.

[0005] To achieve the above object, the technical scheme of the present application is as follows:

[0006] The present application relates to a multi-feature fusion radar target recognition system based on ensemble learning, which comprises:

[0007] An angle-invariant feature extraction network is used to input HRRP radar data at different angles, and a multi-layer sparse autoencoder is used to fuse all single features at different angles to obtain angle-invariant features at different angles; the multi-layer sparse autoencoder comprises an encoder layer, a deep encoding layer, a decoder layer and an output layer, and a regularization constraint is added on the basis of a mean square error function;

[0008] A HRRP radar recognition network equivalent to the angle is used to classify and recognize the angle-invariant features at different angles to form a classification result corresponding to the angle-invariant features at each angle;

[0009] An ensemble decision network is used to generate a recognition result label of the HRRP radar data by means of a voting decision.

[0010] Preferably, the encoder layer is a fully connected encoder layer, which is used to input a time sequence of HRRP radar data sequences in sequence into n neurons of the input layer, and is connected to m neurons of the encoder layer in a fully connected manner, the weight matrix of the encoder layer is W 1 , the dimension of the weight matrix is , the bias vector is b 1 , the dimension of the bias vector is , the neuron activation function is selected as a ReLU activation function, and the output result of the encoder layer is h 1, which is represented as:

[0011] .

[0012] Preferably, the deep encoding layer is used to extract deep features of the HRRP radar data, the number of neurons is l , the weight matrix is W 2 , the dimension of the weight matrix is , the bias vector is b 2 , the dimension is , the neuron activation function is selected as a ReLU activation function, and the output result of the deep encoding layer is h 2, which is represented as:

[0013] .

[0014] Preferably, the decoder layer is used to reconstruct the input HRRP radar data, the decoder layer and the deep encoding layer constitute a symmetric arrangement, the number of neurons of the decoder layer is m , the weight matrix is W 3 , the dimension is , the bias vector is b 3 , the dimension is , the neuron activation function is selected as the ReLU activation function, and the output result of the decoder layer is h 3 is represented as:

[0015] .

[0016] Preferably, the output layer is used to output the angle-invariant features of different angles, the number of nodes of the output layer is w , the weight matrix is W 4 , the dimension is , the bias vector is b 4 , the dimension is , and the output is an angle-invariant feature extraction vector F， is represented as:

[0017] ;

[0018] The loss function of the multi-layer sparse autoencoder is represented as:

[0019] ,

[0020] wherein, J AE is the loss function of the multi-layer sparse autoencoder, W is a matrix composed of the weights of the encoder layer, the deep encoding layer, the decoder layer and the output layer, b is a matrix composed of the bias of the encoder layer, the deep encoding layer, the decoder layer and the output layer, is a regular term sparsity, n is the number of HRRP radar data sequences, X i represents the HRRP radar data sequence numbered i , F i represents the angle-invariant feature extraction vector corresponding to the HRRP radar data sequence numbered i .

[0021] Preferably, each of the HRRP radar recognition networks is a classification recognition network composed of a full connection layer and a softmax classifier added after the full connection layer.

[0022] The application also relates to a multi-feature fusion radar target recognition method based on ensemble learning, which comprises the following steps:

[0023] S1. inputting a HRRP radar data set containing all angles into an angle-invariant feature extraction network, and training the angle-invariant feature extraction network;

[0024] S2. inputting multi-angle target HRRP radar data into the trained angle-invariant feature extraction network, and fusing all single different-angle features by using a multi-layer sparse autoencoder to obtain angle-invariant features of different angles; the multi-layer sparse autoencoder comprises an encoder layer, a deep encoding layer, a decoder layer and an output layer, and a regularization constraint is added on the basis of a mean square error function;

[0025] S3. inputting the angle-invariant features into a HRRP radar recognition network, classifying and recognizing the angle-invariant features of different angles, and forming a classification result corresponding to the angle-invariant features of each angle;

[0026] S4. inputting the classification results of the angle-invariant features of all angles into an ensemble decision network, and generating a recognition result label of the HRRP radar data by means of a voting decision.

[0027] Preferably, the specific steps of S2 for fusing all single different-angle features to obtain angle-invariant features of different angles by using a multi-layer sparse autoencoder are as follows:

[0028] S2.1. inputting a time sequence of HRRP radar data sequences into n neurons of an input layer of the encoder layer in sequence, and connecting the neurons to m neurons of the encoder layer in a full connection manner, wherein a weight matrix of the encoder layer is W 1 , the weight matrix has a dimension of , a bias vector is b 1 , the bias vector has a dimension of , a neuron activation function is selected as a ReLU activation function, and an output result h 1 of the encoder layer is represented as:

[0029] ;

[0030] S2.2. inputting the output result of the encoder layer into the deep encoding layer to extract deep features of the HRRP radar data, wherein the number of neurons of the deep encoding layer is l , a weight matrix is W 2 , the weight matrix has a dimension of The bias vector is b 2 , dimension The neuron activation function is chosen as the ReLU activation function, and the output of the deep coding layer is... h 2 is represented as:

[0031] ;

[0032] S2.3. The output of the deep coding layer is input into the decoder layer to reconstruct the input HRRP radar data. The decoder layer and the deep coding layer are symmetrically configured, and the number of neurons in the decoder layer is [number missing]. m The weight matrix is W 3 , dimension The bias vector is b 3 , dimension The neuron activation function is chosen as the ReLU activation function, and the output of the decoder layer is... h 3 is represented as:

[0033] ;

[0034] S2.4. Input the output of the decoder layer into the output layer. The output layer outputs angular invariant features at different angles. The number of nodes in the output layer is... w There are , and the weight matrix is ​​. W 4 , dimension The bias vector is b 4 , dimension The output is an angle-invariant feature extraction vector. F， Represented as:

[0035] .

[0036] Preferably, the HRRP radar dataset containing all angles in S1 is represented by the following matrix:

[0037] ,

[0038] in, S HRRP Represents the HRRP radar dataset. M Indicates that there is M y represents the angle, where each angle contains y Class target, Y αβ Indicates the first α The first angle β HRRP radar target data for similar targets, and Yαβ ={ X 1, X 2… X t}, α ∈[1~ M ], β ∈[1~y], X t represents a piece of HRRP radar data.

[0039] Preferably, the S4 generates the identification result label of the HRRP radar data by means of voting decision, that is, the final classification label value is obtained by using the weighted average method to integrate the calculation label values of multiple angles, and is represented as:

[0040] ,

[0041] wherein, ω α is the angle α corresponding to the classification weight of the angle classifier under the piece of training data, ω α ≥1 and , Y α represents the label value of the target HRRP radar data with the angle α , X represents the target HRRP radar data, M represents that there are M angles.

[0042] Compared with the prior art, the technical scheme provided by the present application has the following beneficial effects:

[0043] The multi-feature fusion radar target identification system and method based on ensemble learning provided by the present application are aimed at the unique properties of HRRP radar target signal data, and use a multi-layer sparse autoencoder to extract angle-invariant features. Compared with the traditional method, the angle-invariant feature extraction method using the multi-layer sparse autoencoder no longer simply relies on angle features. The method optimizes the loss function of the traditional autoencoder through a sparse regularization term, thereby providing the model with sparsity and certain generalization ability. The multi-classifier based on ensemble learning can integrate multiple features of the target. Through comprehensive analysis of these features, the method can accurately capture the subtle differences between different individuals in the same class of target, thereby effectively avoiding the recognition failure problem caused by individual differences. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is a model schematic diagram of the multi-feature fusion radar target identification method based on ensemble learning of the present application;

[0045] Figure 2 The structure schematic diagram of the multi-layer sparse autoencoder of the present application;

[0046] Figure 3 The structure schematic diagram of the classifier of the present application;

[0047] Figure 4 The structure schematic diagram of the multi-classifier containing integrated decision of the present application. DETAILED DESCRIPTION

[0048] For further understanding of the present application, the present application is described in detail in conjunction with the embodiments, the following embodiments are used to illustrate the present application, but not to limit the scope of the present application.

[0049] Refer to the drawings Figure 1 The present application relates to a multi-feature fusion radar target recognition system based on integrated learning, which comprises an angle invariant feature extraction network, an HRRP radar recognition network with equal angle and an integrated decision network.

[0050] The angle invariant feature extraction network is used for inputting HRRP radar data at different angles, and a multi-layer sparse autoencoder is used to fuse all single features at different angles to obtain angle invariant features at different angles; the multi-layer sparse autoencoder comprises an encoder layer, a deep encoding layer, a decoder layer and an output layer, and a regularization constraint is added on the basis of mean square error function, as shown in Figure 2

[0051] The encoder layer is a fully connected encoder layer, which is used for inputting time sequence HRRP radar data sequence in order into n neurons of the input layer, and connected to m neurons of the encoder layer in a fully connected manner, the weight matrix of the encoder layer is W 1 , the dimension of the weight matrix is , the bias vector is b 1 , the dimension of the bias vector is , the neuron activation function is ReLU activation function, and the output result of the encoder layer is h 1, which is:

[0052]

[0053] In Figure 2 , the output of this layer is h 1 , i.e. the neuron X 1 ~X n of this layer.​​

[0054] The deep encoding layer is used to extract deep features of the HRRP radar data, the number of neurons is l , the weight matrix is W 2 , the dimension of the weight matrix is , the bias vector is b 2 , the dimension is , the neuron activation function selects the ReLU activation function, and the output result of the deep encoding layer is h 2, which is expressed as:

[0055] ;

[0056] In the Figure 2 , h 2 , the output of the neurons is K 1 ~K m of this layer;

[0057] The decoder layer is used to reconstruct the input HRRP radar data, and the decoder layer and the deep encoding layer are symmetrically arranged, the number of neurons of the decoder layer is m , the weight matrix is W 3 , the dimension is , the bias vector is b 3 , the dimension is , the neuron activation function selects the ReLU activation function, and the output result of the decoder layer is h 3, which is expressed as:

[0058] ;

[0059] In the Figure 2 , h 3 , the output of the neurons is H 1 ~H m of this layer;

[0060] The output layer is used to output angle-invariant features of different angles, the number of nodes of the output layer is w , the weight matrix is W 4 , the dimension is , the bias vector is b 4 , the dimension is , and the output is an angle-invariant feature extraction vector F， , which is expressed as:

[0061] ;

[0062] The loss function of the autoencoder at this time can be represented by a mean square error function, as shown in the following formula:

[0063] ,

[0064] On the basis of the loss function, a regularization constraint term is further added to improve the generalization ability of the original autoencoder and bring sparsity to the autoencoder. At this time, the loss function of the multi-layer sparse autoencoder is represented as:

[0065] ,

[0066] Wherein, J AE is the loss function of the multi-layer sparse autoencoder, W is a matrix composed of the weights of the encoder layer, the deep encoding layer, the decoder layer and the output layer, b is a matrix composed of the biases of the encoder layer, the deep encoding layer, the decoder layer and the output layer, is a regularization term sparsity, n is the number of HRRP radar data sequences, X i represents the HRRP radar data sequence numbered i , F i represents the HRRP radar data sequence numbered i corresponding to the angle invariant feature extraction vector.

[0067] Each of the different angle HRRP radar identification networks is as shown in Figure 3 and Figure 4 , specifically, a layer of fully connected layer is added after a layer of softmax classifier to form a classification identification network, and the obtained feature information is classified and identified as a specific classification label value, which is used for classifying and identifying the input different angle angle invariant feature F to form the classification result corresponding to the angle invariant feature of each angle.

[0068] The integrated decision network is used to generate the identification result label of the HRRP radar data by means of voting decision.

[0069] The multi-feature fusion radar target identification method based on ensemble learning comprises the following steps:

[0070] S1. Input the HRRP radar data set containing all angles into the angle invariant feature extraction network, and train the angle invariant feature extraction network. The HRRP radar data set is represented by a matrix:

[0071] ,

[0072] in, S HRRP Represents the HRRP radar dataset. M Indicates that there is M y represents the angle, where each angle contains y Class target, Y αβ Indicates the first α The first angle β HRRP radar target data for similar targets, and Y αβ ={ X 1, X 2… X t}, α ∈[1~ M ], β ∈[1~y], X t This represents a single HRRP radar data point;

[0073] HRRP radar data X t For time-domain data, represented as , x n Represents radar time-domain data sequences;

[0074] During training, the aforementioned HRRP radar data will be used. X t The inputs are sequentially fed into the multi-feature fusion radar target recognition method based on ensemble learning, and the angle-invariant feature extraction network, HRRP radar recognition network and ensemble decision network are trained in the manner of S2~S4. The training process will not be described in detail.

[0075] S2. See Appendix Figure 1 and attached Figure 4 As shown, during the recognition process, multi-angle target HRRP radar data is input into a trained angle-invariant feature extraction network. A multi-layer sparse autoencoder is used to fuse all single features from different angles to obtain angle-invariant features at different angles. The multi-layer sparse autoencoder includes an encoder layer, a deep encoder layer, a decoder layer, and an output layer, and a regularization constraint is added to the mean square error function. The specific steps are as follows:

[0076] S2.1. Input the time-series HRRP radar data sequence sequentially into the input layer of the encoder layer. n Within each neuron, and connected in a fully connected manner to the encoder layer.m The neurons are connected, the weight matrix of the encoder layer is W 1 The dimension of the weight matrix is The bias vector is b 1 The dimension of the bias vector is The neuron activation function selects the ReLU activation function, and the output result of the encoder layer is h 1, which is expressed as:

[0077] ;

[0078] S2.2. The output result of the encoder layer is input into the deep encoding layer to extract the deep features of the HRRP radar data, the number of neurons of the deep encoding layer is l , the weight matrix is W 2 The dimension of the weight matrix is The bias vector is b 2 The dimension is The neuron activation function selects the ReLU activation function, and the output result of the deep encoding layer is h 2, which is expressed as:

[0079] ;

[0080] S2.3. The output result of the deep encoding layer is input into the decoder layer to reconstruct the input HRRP radar data, the decoder layer and the deep encoding layer are symmetrically arranged, the number of neurons of the decoder layer is m , the weight matrix is W 3 The dimension is The bias vector is b 3 The dimension is The neuron activation function selects the ReLU activation function, and the output result of the decoder layer is h 3, which is expressed as:

[0081] ;

[0082] S2.4. The output result of the decoder layer is input into the output layer, and the output layer outputs the angle-invariant features at different angles, the number of nodes of the output layer is w , the weight matrix is W 4 The dimension is The bias vector is b 4 The dimension is The output is an angle-invariant feature extraction vector F， , which is expressed as:

[0083]

[0084] S3. Input the angle-invariant features into the HRRP radar recognition network, classify and recognize the angle-invariant features of different angles, and form a classification result corresponding to the angle-invariant features of each angle;

[0085] S4. Input the classification results of the angle-invariant features of all angles into the integrated decision network, generate the recognition result label of the HRRP radar data through the voting decision method, specifically, use the weighted average method for integrated decision, that is, obtain the final classification label value through the weighted average method from the calculation label values of multiple angles, and the final classification label value is represented as:

[0086] ,

[0087] wherein, ω α is the classification weight of each angle classifier under the training data, ω α ≥1 and , Y α represents the label value of the target HRRP radar data with the angle α X represents the target HRRP radar data, M represents that there are M angles.

[0088] The application is described in detail in combination with the embodiments, but the content described is only the preferred embodiments of the application and cannot be considered as limiting the implementation scope of the application. Any equivalent changes and improvements made according to the application scope should still belong to the patent coverage range of the application.

Claims

1. A multi-feature fusion radar target recognition system based on ensemble learning, characterized in that: It includes: An angle-invariant feature extraction network is used to input HRRP radar data at different angles, and a multi-layer sparse autoencoder is used to fuse all the individual features at different angles to obtain angle-invariant features at different angles. The multilayer sparse autoencoder includes an encoder layer, a deep encoder layer, a decoder layer, and an output layer, and adds regularization constraints to the mean square error function. The encoder layer is a fully connected encoder layer used to sequentially input the time-series HRRP radar data sequences into the input layer. n In one neuron, and connected in a fully connected manner to the encoder layer m The encoder layer has neurons connected together, and its weight matrix is ​​. W 1 The dimension of the weight matrix is m × n 1. The bias vector is b 1 The dimension of the bias vector is The neuron activation function is chosen as the ReLU activation function, and the output of the encoder layer is... h 1 is represented as: ； The deep coding layer is used to extract deep features from HRRP radar data, and its number of neurons is [number missing]. l The weight matrix is W 2 The weight matrix has dimensions of The bias vector is b 2 , dimension The neuron activation function is chosen as the ReLU activation function, and the output of the deep coding layer is... h 2 is represented as: ； The decoder layer is used to reconstruct the input HRRP radar data. The decoder layer and the deep coding layer are symmetrically configured, and the number of neurons in the decoder layer is [number missing]. m The weight matrix is W 3 , dimension The bias vector is b 3 , dimension The neuron activation function is chosen as the ReLU activation function, and the output of the decoder layer is... h 3 is represented as: ； The output layer is used to output angular invariant features at different angles, and the number of nodes in the output layer is [number missing]. w There are , and the weight matrix is ​​. W 4 , dimension The bias vector is b 4 , dimension The output is an angle-invariant feature extraction vector. F， Represented as: ； The loss function of the multilayer sparse autoencoder is expressed as: ， in, J AE The loss function of a multilayer sparse autoencoder is... W It is a matrix composed of the weights of the encoder layer, deep encoder layer, decoder layer, and output layer. b It is a matrix composed of the biases of the encoder layer, deep coding layer, decoder layer, and output layer. For regularization terms to be sparse, n For HRRP radar data sequence number, X i Indicates the number is i HRRP radar data sequence, F i Indicates the number is i The angle-invariant feature extraction vector corresponding to the HRRP radar data sequence; An HRRP radar identification network with the same amount of angle is used to classify and identify angle-invariant features at different angles, forming a classification result corresponding to the angle-invariant features at each angle; An integrated decision network is used to generate identification result labels for HRRP radar data through a voting decision-making process.

2. The multi-feature fusion radar target recognition system based on ensemble learning according to claim 1, characterized in that: Each of the HRRP radar recognition networks is a classification and recognition network consisting of a fully connected layer followed by a softmax classifier.

3. A multi-feature fusion radar target recognition method based on ensemble learning, characterized in that, It includes the following steps: S1. Input the HRRP radar dataset containing all angles into the angle-invariant feature extraction network and train the angle-invariant feature extraction network; S2. Input the multi-angle target HRRP radar data into the trained angle-invariant feature extraction network, and use a multi-layer sparse autoencoder to fuse all the single features from different angles to obtain angle-invariant features from different angles; the loss function of the multi-layer sparse autoencoder is expressed as: ， in, J AE The loss function of a multilayer sparse autoencoder is... W It is a matrix composed of the weights of the encoder layer, deep encoder layer, decoder layer, and output layer. b It is a matrix composed of the biases of the encoder layer, deep coding layer, decoder layer, and output layer. For regularization terms to be sparse, n For HRRP radar data sequence number, X i Indicates the number is i HRRP radar data sequence, F i Indicates the number is i The angle-invariant feature extraction vector corresponding to the HRRP radar data sequence; The specific steps to obtain the angle-invariant features at different angles are as follows: S2.

1. Input the time-series HRRP radar data sequence sequentially into the input layer of the encoder layer. n In one neuron, and connected in a fully connected manner to the encoder layer m The encoder layer has neurons connected together, and its weight matrix is ​​. W 1 The dimension of the weight matrix is m × n 1. The bias vector is b 1 The dimension of the bias vector is The neuron activation function is chosen as the ReLU activation function, and the output of the encoder layer is... h 1 is represented as: ； S2.

2. The output of the encoder layer is input into the deep coding layer to extract deep features from the HRRP radar data. The number of neurons in the deep coding layer is... l The weight matrix is W 2 The weight matrix has dimensions of The bias vector is b 2 , dimension The neuron activation function is chosen as the ReLU activation function, and the output of the deep coding layer is... h 2 is represented as: ； S2.

3. The output of the deep coding layer is input into the decoder layer to reconstruct the input HRRP radar data. The decoder layer and the deep coding layer are symmetrically configured, and the number of neurons in the decoder layer is [number missing]. m The weight matrix is W 3 , dimension The bias vector is b 3 , dimension The neuron activation function is chosen as the ReLU activation function, and the output of the decoder layer is... h 3 is represented as: ； S2.

4. Input the output of the decoder layer into the output layer. The output layer outputs angular invariant features at different angles. The number of nodes in the output layer is... w There are , and the weight matrix is ​​. W 4 , dimension The bias vector is b 4 , dimension The output is an angle-invariant feature extraction vector. F， Represented as: ； S3. Input the angle-invariant features into the HRRP radar recognition network, classify and recognize the angle-invariant features at different angles, and form a classification result corresponding to the angle-invariant features at each angle; S4. Input the classification results of all angle-invariant features into the ensemble decision network, and generate the identification result labels for the HRRP radar data through a voting decision method. That is, use a weighted average method for ensemble decision-making to obtain the final classification label value, represented as: ， in, ω α For angle α The corresponding classification weights of the angle classifier in the target HRRP radar data, ω α ≤1 and , Y α The angle is represented as α The tag value of the target HRRP radar data. X This indicates target HRRP radar data. M Indicates that there is M From one angle.

4. The multi-feature fusion radar target recognition method based on ensemble learning according to claim 3, characterized in that: The HRRP radar dataset containing all angles in S1 is represented by the following matrix: ， in, S HRRP Represents the HRRP radar dataset. M Indicates that there is M y represents the angle, where each angle contains y Class target, Y αβ Indicates the first α The first angle β HRRP radar target data for similar targets, and Y αβ ={ X 1, X 2… X t }, α ∈[1~ M ], β ∈[1~y], X t Indicates the number is t HRRP radar data.

Images