A method for air target recognition based on micro-motion modulation and convolutional neural network

Abstract

The present invention discloses a method for identifying air targets based on micro-modulation and convolutional neural networks. This method introduces convolutional neural networks into radar target recognition and is primarily applicable to the classification and identification of jet aircraft, propeller aircraft, and helicopters by coherent radars operating in a high-repetition-rate mode. The main process is as follows: first, data preprocessing is performed on the Doppler micro-modulation frequency domain data; then, the preprocessed Doppler micro-modulation feature frequency domain data is fed into a convolutional neural network for feature extraction and classification and identification. By constructing a convolutional neural network model suitable for processing micro-Doppler features, the present invention fully exploits the deep-level attribute characteristics of the target contained in the target micro-Doppler frequency domain data, achieves robust automatic feature extraction, and completes the classification of three types of aircraft targets.

Description

Technical Field

[0001] The invention belongs to the technical field of radar target recognition. Background Art

[0002] The target's micro-Doppler modulation effect, which contains detailed information about the geometry and micro-motion of its micro-moving components, offers a new approach to automatic target recognition (ATR). In recent years, research on the micro-Doppler effect has yielded impressive results. For the classification and identification of dynamic aircraft targets, the differences in micro-motion component modulation can also be used to distinguish between different types of aircraft targets. The micro-moving component structures of the three types of aircraft targets (helicopters, propeller aircraft, and jets) vary significantly, such as rotor length, rotational speed, and number, resulting in distinct micro-Doppler modulation characteristics in their narrowband radar echoes.

[0003] In recent years, a wealth of research has been conducted on using differences in micromotion component characteristics to distinguish different targets. This work has primarily focused on two areas: physics-driven classification feature extraction and data-driven classification feature extraction. Feature extraction is fundamental and crucial for both physics-driven and data-driven approaches. Consequently, many researchers have devoted significant attention to exploring various feature extraction methods. For example, in the paper "Aircraft Target Classification Based on Empirical Mode Decomposition," empirical mode decomposition (EMD)-based feature extraction was used to identify three types of aircraft targets based on their Doppler frequency modulation, combined with support vector machines (SVMs) for target classification. In the field of feature learning, deep networks possess powerful nonlinear feature extraction capabilities, and related theoretical and experimental results continue to emerge. In 2006, Hinton et al. proposed an unsupervised, layer-by-layer greedy training method, addressing the "gradient dissipation" problem associated with increased depth and enabling the development of deeper neural networks. Subsequently, numerous researchers have proposed various DL models, such as deep belief networks (DBNs) and convolutional neural networks (CNNs), tailored to different application contexts. These models have achieved superior results compared to traditional feature extraction methods in image processing and other fields.

[0004] Different from the methods proposed in other literatures, this paper applies convolutional neural networks to the field of micro-Doppler target recognition based on micro-Doppler frequency domain features, reduces manual operations in the feature selection process, fully explores the deep-level attribute characteristics of the target contained in the target micro-Doppler frequency domain data, realizes robust automatic feature extraction, and completes the classification of three types of aircraft targets. Summary of the Invention

[0005] The purpose of the present invention is to provide an air target recognition method based on micro-modulation and convolutional neural network, which introduces convolutional neural network into radar target recognition. It is mainly suitable for the classification and recognition of jet aircraft, propeller aircraft and helicopters by coherent radar with high repetition rate working mode.

[0006] In order to achieve the above technical objectives, the technical solutions of the present invention are:

[0007] First, the fixed clutter of the Doppler micro-modulation spectrum data is removed using the same-frequency difference method; the fuselage frequency component in the Doppler micro-modulation spectrum data after the fixed clutter is removed is shifted to zero frequency; the Doppler micro-modulation frequency domain data after the fuselage frequency component is shifted to zero frequency is normalized using the L2 regularization method; then, a convolutional neural network with 6 convolutional layers and 3 pooling layers is established to extract deep attribute features from the pre-processed Doppler micro-modulation frequency domain data through the convolutional and pooling layers; finally, two fully connected layers and a softmax classifier are established to classify and identify the extracted deep attribute features of the target. The specific steps include:

[0008] Step (1): obtaining pre-processed Doppler micro-modulation frequency domain data;

[0009] Step (2): Obtain the target deep-level attribute features and classification recognition results extracted by the convolutional neural network.

[0010] Wherein step (1) further comprises:

[0011] Step A uses the same-frequency difference method to remove the fixed clutter of the Doppler micro-modulation frequency domain data;

[0012] Step B: shifting the fuselage frequency component in the Doppler micro-modulation frequency domain data after removing the fixed clutter to zero frequency;

[0013] Step C uses the L2 regularization method to normalize the data obtained in step B.

[0014] Step (2) further includes:

[0015] Step D: Establish a convolutional neural network with 6 convolutional layers and 3 pooling layers, and extract the deep attribute features of the target from the Doppler micro-modulation frequency domain data obtained in step (1) through the convolutional layers and pooling layers;

[0016] In step E, two fully connected layers and a softmax classifier are established to classify and identify the extracted deep attribute features of the target.

[0017] Compared with the prior art, the present invention has the following significant advantages:

[0018] A method for removing fixed clutter from Doppler micro-motion modulation frequency domain data can quickly and effectively detect the fixed clutter frequency range. This method features good adaptability, low computational complexity, and high operational efficiency. A convolutional neural network model and loss function construction method effectively prevent the network from overfitting, fully exploiting the deep-level target attribute characteristics contained in the target micro-Doppler micro-motion frequency domain data, and exhibiting excellent recognition results and generalization capabilities. This invention has high application value in the field of radar target recognition.

[0019] The present invention is further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a data processing flow chart of the present invention.

[0021] Figure 2 This is a schematic diagram of the Doppler micro-modulation frequency domain data of the present invention, where 1 is the fuselage frequency component, 2 is the fixed clutter frequency component, and 3 is the micro-Doppler frequency component.

[0022] Figure 3 1 is a schematic diagram of Doppler micro-modulation frequency domain data after eliminating fixed clutter in the present invention, wherein 1 is the fuselage frequency component and 3 is the micro-Doppler frequency component.

[0023] Figure 4 1 is a schematic diagram of Doppler micro-modulation frequency domain data after the fuselage frequency component is shifted to zero frequency and normalized according to the present invention, wherein 1 is the fuselage frequency component and 3 is the micro-Doppler frequency component.

[0024] Figure 5 It is a schematic diagram of the convolutional neural network structure of the present invention. DETAILED DESCRIPTION

[0025] The present invention provides an air target recognition method based on micro-modulation and convolutional neural network, which effectively realizes the classification and recognition of jet aircraft, propeller aircraft and helicopters in the air by radar. Through the present invention, it is possible to fully exploit the deep attribute characteristics of the target contained in the target micro-Doppler frequency domain data, realize robust automatic feature extraction, and complete the classification of the three types of aircraft targets. For the specific implementation process, please refer to the attached Figure 1 , the preferred embodiment steps are as follows:

[0026] Step (1) obtains the pre-processed Doppler micro-modulation frequency domain data as follows:

[0027] Step A: Use the frequency difference method to remove the fixed clutter of Doppler micro-modulation frequency domain data, as shown in the attached figure. Figure 3 , specifically:

[0028] 1) Obtain the spectrum p0 corresponding to the target distance unit, with the number of spectrum sampling points being 256.

[0029] 2) Taking the target distance unit as the center, take 5 distance units on the left and right, and obtain the spectrum of the 5 distance units on the left and right respectively, which is recorded as

[0030] P={p i}, i=-5,-4,-3,-2,-1,1,2,3,4,5

[0031] where p i is the spectrum of the i-th distance unit.

[0032] 3) Calculate the frequency variant kurtosis set K of each spectrum = {k i}:

[0033]

[0034] where p i is the spectrum of the i-th distance unit, μ i is the spectrum mean, σ i is the standard deviation of the spectrum, E(p i -μ i ) 4 is the fourth-order center distance of the spectrum of the i-th distance unit, E(p i -μ i ) 3 is the third-order center distance of the spectrum of the i-th distance unit.

[0035] 4) Find the maximum element max({k i}) The corresponding distance unit spectrum is denoted as p k .

[0036] 5) Take the target distance unit spectrum p0 and p k As the basis, calculate the spectrum difference p′,

[0037] p′=p0-p k

[0038] 6) In the spectrum difference p′, extend the search to the left and right with zero frequency as the center, and stop the extended search when the following conditions are met:

[0039]

[0040] where p′ r is the frequency value corresponding to the rth frequency point when extending the search to the right, p′ l It is the frequency value corresponding to the lth frequency point when extending the search to the left.

[0041] 7) With p kThe frequency value between the rth frequency point and the lth frequency point replaces the frequency value between the rth frequency point and the lth frequency point of the target range unit spectrum p0 to obtain the Doppler micro-modulation frequency domain data p'' after eliminating the fixed clutter.

[0042] Step B: Shift the fuselage frequency component in the Doppler micro-modulation frequency domain data after removing the fixed clutter to zero frequency.

[0043] Step C: Use L2 regularization method to normalize the data obtained in step B, as shown in the attached figure. Figure 4 .

[0044] Step (2) is to obtain the target deep-level attribute features and classification recognition results extracted by the convolutional neural network, as follows:

[0045] Step D: Establish a convolutional neural network with 6 convolution layers and 3 pooling layers. The Doppler micro-modulation frequency domain data obtained in step (1) is used to extract the deep attribute features of the target through the convolution layer and pooling layer. Figure 5 , specifically:

[0046] 1) Perform the first layer of convolution processing on the pre-processed Doppler micro-modulation frequency domain data obtained in step (1) to obtain the first layer feature map H1.

[0047]

[0048] Where H0 represents the pre-processed Doppler micro-modulation frequency domain data obtained in step (1), W1 represents the weight vector of the first layer convolution kernel, the operator * represents the convolution operation, and b1 represents the offset vector of the first layer. Represents the Relu activation function.

[0049] 2) Perform the second layer convolution processing to obtain the second layer feature map H2,

[0050]

[0051] Where W2 represents the weight vector of the convolution kernel of the second layer, the operator * represents the convolution operation, and b2 represents the offset vector of the second layer. Represents the Relu activation function.

[0052] 3) Perform the third layer pooling process to obtain the third layer feature map H3,

[0053] H3=subsampling(H2)

[0054] Where subsampling represents maximum pooling.

[0055] 4) Perform convolution processing on the fourth layer to obtain the fourth layer feature map H4,

[0056]

[0057] Where W4 represents the weight vector of the 4th layer convolution kernel, the operator * represents the convolution operation, and b4 represents the offset vector of the 4th layer. Represents the Relu activation function.

[0058] 5) Perform convolution processing on the fifth layer to obtain the fifth layer feature map H5.

[0059]

[0060] Where W5 represents the weight vector of the 5th layer convolution kernel, the operator * represents the convolution operation, and b5 represents the offset vector of the 5th layer. Represents the Relu activation function.

[0061] 6) Perform the 6th layer pooling process to obtain the 6th layer feature map H6,

[0062] H6=subsampling(H5)

[0063] Where subsampling represents maximum pooling.

[0064] 7) Perform convolution processing on the 7th layer to obtain the 7th layer feature map H7,

[0065]

[0066] Where W7 represents the weight vector of the convolution kernel of the 7th layer, the operator * represents the convolution operation, and b7 represents the offset vector of the 7th layer. Represents the Relu activation function.

[0067] 8) Perform convolution processing on the 8th layer to obtain the 8th layer feature map H8,

[0068]

[0069] Where w8 represents the weight vector of the 8th layer convolution kernel, the operator * represents the convolution operation, and b8 represents the offset vector of the 8th layer. Represents the Relu activation function.

[0070] 9) Perform the 9th layer pooling process to obtain the 9th layer feature map H9,

[0071] H9=subsampling(H8)

[0072] Where subsampling represents maximum pooling.

[0073] Step E: Establish two fully connected layers and softmax classifier to classify and identify the extracted target deep attribute features, as shown in the attached figure. Figure 5 , specifically:

[0074] 1) Perform the first fully connected layer processing to obtain the value of the first FC block, which is expressed as follows:

[0075]

[0076] Where W 10 represents the weight vector of the first fully connected layer, the operator · represents the dot product operation, b 10 represents the offset vector of the first fully connected layer, Represents the Relu activation function.

[0077] 2) Perform the second fully connected layer processing to obtain the value of the second FC block, which is expressed as follows:

[0078]

[0079] Where W 11 represents the weight vector of the second fully connected layer, the operator · represents the dot product operation, b 12 represents the offset vector of the second fully connected layer, Represents the Relu activation function.

[0080] 3) Use the softmax classifier to classify jets, propeller planes, and helicopters.

[0081] In the process of extracting deep attribute features and classifying targets using convolutional neural networks, the cross entropy loss function and L1 regularization are used to establish the loss function of the convolutional neural network, which can be expressed as follows:

[0082]

[0083] where y i is the label of the correct solution, p i is the probability distribution output by the softmax function, and α is the coefficient.

Claims

1. A method for identifying empty targets based on micro-motion modulation and convolutional neural networks, characterized by: Step (1): obtaining pre-processed Doppler micro-modulation frequency domain data; Step (2): Obtain the target deep-level attribute features and classification recognition results extracted by the convolutional neural network; Wherein said step (1) comprises: Step A: Use the same-frequency difference method to remove the fixed clutter of Doppler micro-modulation frequency domain data. The implementation method is as follows: 1) Obtain the spectrum p0 corresponding to the target distance unit, with the number of spectrum sampling points being 256; 2) Taking the target distance unit as the center, take 5 distance units on the left and right, and obtain the spectrum of the 5 distance units on the left and right respectively, which is recorded as: P = {p i },i=-5,-4,-3,-2,-1,1,2,3,4,5; where p i is the spectrum of the i-th distance unit; 3) Calculate the frequency variant kurtosis set K of each spectrum = {k i }: where p i is the spectrum of the i-th distance unit, μ i is the spectrum mean, σ i is the standard deviation of the spectrum, E(p i -μ i ) 4 is the fourth-order center distance of the spectrum of the i-th distance unit, E(p i -μ i ) 3 is the third-order center distance of the spectrum of the i-th distance unit; 4) Find the maximum element max({k i }) The corresponding distance unit spectrum is denoted as p k ; 5) Take the target distance unit spectrum p0 and p k As the basis, calculate the spectrum difference p′: p′=p0-p k 6) In the spectrum difference p′, extend the search to the left and right with zero frequency as the center, and stop the extended search when the following conditions are met: where p′ r is the frequency value corresponding to the rth frequency point when the search is extended to the right, and p′1 is the frequency value corresponding to the lth frequency point when the search is extended to the left; 7) With p k The frequency value between the rth frequency point and the lth frequency point replaces the frequency value between the rth frequency point and the lth frequency point of the target range unit spectrum p0 to obtain the Doppler micro-modulation frequency domain data p'' after eliminating the fixed clutter; Step B: Shift the fuselage frequency component in the Doppler micro-modulation frequency domain data after removing the fixed clutter to zero frequency; Step C: Normalize the data obtained in step B using the L2 regularization method; The step (2) comprises: Step D: Establish a convolutional neural network with 6 convolutional layers and 3 pooling layers, and extract the deep attribute features of the target from the Doppler micro-modulation frequency domain data obtained in step (1) through the convolutional layers and pooling layers; Step E: Establish two fully connected layers and a softmax classifier to classify and identify the extracted deep attribute features of the target.

2. The method for identifying an empty target based on micro-motion modulation and convolutional neural network according to claim 1, characterized in that: The step D comprises: 1) Perform the first layer of convolution processing on the pre-processed Doppler micro-modulation frequency domain data obtained in step (1) to obtain the first layer feature map H1: Where H0 represents the pre-processed Doppler micro-modulation frequency domain data obtained in step (1), W1 represents the weight vector of the first layer convolution kernel, the operator * represents the convolution operation, and b1 represents the offset vector of the first layer. Represents the Relu activation function; 2) Perform the second layer of convolution processing to obtain the second layer feature map H2: Where W2 represents the weight vector of the convolution kernel of the second layer, the operator * represents the convolution operation, and b2 represents the offset vector of the second layer. Represents the Relu activation function; 3) Perform the third layer pooling process to obtain the third layer feature map H3: H3=subsampling(H2) Where subsampling means maximum pooling; 4) Perform convolution processing on the fourth layer to obtain the fourth layer feature map H4: Where W4 represents the weight vector of the 4th layer convolution kernel, the operator * represents the convolution operation, and b4 represents the offset vector of the 4th layer. Represents the Relu activation function; 5) Perform convolution processing on the fifth layer to obtain the fifth layer feature map H5: Where W5 represents the weight vector of the 5th layer convolution kernel, the operator * represents the convolution operation, and b5 represents the offset vector of the 5th layer. Represents the Relu activation function; 6) Perform the sixth layer pooling process to obtain the sixth layer feature map H6: H6=subsampling(H5) Where subsampling means maximum pooling; 7) Perform convolution processing on the 7th layer to obtain the 7th layer feature map H7: Where W7 represents the weight vector of the convolution kernel of the 7th layer, the operator * represents the convolution operation, and b7 represents the offset vector of the 7th layer. Represents the Relu activation function; 8) Perform convolution processing on the 8th layer to obtain the 8th layer feature map H8: Where W8 represents the weight vector of the 8th layer convolution kernel, the operator * represents the convolution operation, and b8 represents the offset vector of the 8th layer. Represents the Relu activation function; 9) Perform the 9th layer pooling process to obtain the 9th layer feature map H9: H9=subsampling(H8) Where subsampling represents maximum pooling.

3. The method for identifying empty targets based on micro-motion modulation and convolutional neural network according to claim 1, characterized in that: The step E comprises: 1) Perform the first fully connected layer processing to obtain the value of the first FC block, which is expressed as follows: Where W 10 represents the weight vector of the first fully connected layer, the operator · represents the dot product operation, b 10 represents the offset vector of the first fully connected layer, Represents the Relu activation function; 2) Perform the second fully connected layer processing to obtain the value of the second FC block, which is expressed as follows: Where W 11 represents the weight vector of the second fully connected layer, the operator · represents the dot product operation, b 12 represents the offset vector of the second fully connected layer, Represents the Relu activation function; 3) Use the softmax classifier to classify jets, propeller planes, and helicopters.

4. The method for identifying empty targets based on micro-motion modulation and convolutional neural networks according to claim 2 or claim 3, characterized in that: The loss function established by the convolutional neural network includes: using the cross entropy loss function and L1 regularization to establish the loss function of the present invention, which is expressed by the following formula: where y i is the label of the correct solution, p i is the probability distribution output by the softmax function, and α is the coefficient.

Images