Rotorcraft target recognition method based on micro-Doppler characteristics and deep learning

By combining FEKO simulation and MATLAB theoretical calculations with a ResNet18-SVM neural network based on short-time Fourier transform and attention mechanism, the problems of inaccurate acquisition and low recognition rate of UAV radar echo signals were solved, achieving low-cost and high-accuracy rotorcraft target recognition.

CN118566873BActive Publication Date: 2024-11-01ANHUI UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411030896.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-30
Publication Date
2024-11-01
Estimated Expiration
2044-07-30

AI Technical Summary

Technical Problem

Existing technologies are costly and inaccurate in acquiring radar echo signals from drones, while deep learning networks have low accuracy in micro-Doppler feature recognition, making it difficult to effectively identify rotorcraft targets.

Method used

The radar echo signal of the UAV was obtained through FEKO simulation and MATLAB theoretical calculation. Micro-Doppler features were extracted by combining short-time Fourier transform and classified and identified using a ResNet18-SVM neural network based on attention mechanism.

Benefits of technology

It achieves low-cost and accurate acquisition of UAV radar echoes, improving the accuracy and robustness of target recognition for rotorcraft, especially in complex backgrounds.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118566873B_ABST
    Figure CN118566873B_ABST
Patent Text Reader

Abstract

The present invention relates to a method for identifying rotorcraft targets based on micro-Doppler features and deep learning, including: simulating the echo signals of a single-rotor four-blade UAV, a single-rotor three-blade UAV, and a four-rotor UAV respectively; obtaining the echo signals of a single-rotor four-blade UAV, a single-rotor three-blade UAV, and a four-rotor UAV respectively through theoretical calculation; performing time-frequency transformation to obtain a time-frequency graph data set; dividing into a training set and a test set; constructing and training a ResNet18-SVM neural network based on an attention mechanism; classifying the time-frequency graphs of three types of rotorcraft, and realizing target identification of the three types of rotorcraft. The present invention can replace actual measurement, and obtain more accurate drone radar echoes more conveniently and at a lower cost; the ResNet18-SVM neural network based on the attention mechanism of the present invention has good performance in target identification based on aircraft micro-Doppler features.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of radar technology, signal processing and deep learning technology, and in particular to a rotorcraft target recognition method based on micro-Doppler features and deep learning. Background Art

[0002] Drones and other low-speed, slow, and small aircraft are widely used in life, military, and other fields due to their lower operating difficulty and lower purchase cost. The wide application leads to the leakage of privacy information, posing a huge challenge to social public security. Different shapes and rotation speeds of drones will produce different micro-Doppler features. Based on this, micro-Doppler feature extraction can realize drone target recognition.

[0003] At present, there are three main ways to obtain drone radar echoes containing micro-Doppler features: actual measurement, simulation, and theoretical calculation. Among them, actual measurement is costly and it is difficult to obtain a rich data set, while simulation and theoretical calculation do not take into account the different azimuth, pitch angle, and position of each rotor relative to the radar emission source, as well as the different rotation speeds of each rotor, so the obtained drone radar echo signal is not accurate enough. The existing deep learning network model does not perform well in the problem of aircraft target recognition based on micro-Doppler features, and the recognition accuracy needs to be improved. Summary of the invention

[0004] In order to solve the problems that the UAV radar echo signal obtained by the prior art is not accurate enough and the existing network does not perform well in the problem of aircraft target recognition based on micro-Doppler features, the purpose of the present invention is to provide a rotorcraft target recognition method based on micro-Doppler features and deep learning, which can extract more accurate radar echo signals and improve the classification accuracy of micro-Doppler time-frequency diagrams.

[0005] To achieve the above object, the present invention adopts the following technical solution: a method for rotorcraft target recognition based on micro-Doppler characteristics and deep learning, the method comprising the following steps in order:

[0006] (1) Simulate the echo signals of a single-rotor four-blade UAV, a single-rotor three-blade UAV, and a four-rotor UAV respectively;

[0007] (2) The echo signals of a single-rotor four-blade UAV, a single-rotor three-blade UAV, and a four-rotor UAV are obtained through theoretical calculations;

[0008] (3) performing time-frequency transformation on the echo signals described in steps (1) and (2) by short-time Fourier transform, and obtaining a time-frequency graph data set representing micro-Doppler characteristics of a single-rotor four-blade UAV, a single-rotor three-blade UAV, and a quad-rotor UAV;

[0009] (4) Divide the time-frequency graph dataset into a training set and a test set in a ratio of 8:2;

[0010] (5) Construct a ResNet18-SVM neural network based on the attention mechanism;

[0011] (6) Train the ResNet18-SVM neural network based on the attention mechanism through the training set to obtain the trained ResNet18-SVM neural network based on the attention mechanism, and test the ResNet18-SVM neural network based on the attention mechanism through the test set;

[0012] (7) Use the trained ResNet18-SVM neural network based on the attention mechanism to classify the time-frequency graphs of the three types of rotorcraft and realize target recognition of the three types of rotorcraft.

[0013] The step (1) specifically includes: establishing a single-rotor four-blade UAV model and a single-rotor three-blade UAV model and importing them into FEKO software, placing the rotor center at the origin of the coordinate system, setting the incident wave frequency, setting the pitch angle to an arbitrary fixed value, setting the far field to a single station, setting the plane wave azimuth angle to start from a certain initial angle, uniformly sampling to a range of the initial angle plus 360°, simulating the rotation of the rotor with the change of the plane wave azimuth angle, using the grid search function to change the initial angle and pitch angle of the plane wave azimuth angle, running the FEKO software to obtain a .ffe file sequence, processing each .ffe file, and extracting echo data;

[0014] For a quadrotor drone, a quadrotor drone model is established and imported into the FEKO software, the incident wave frequency is set, the plane wave incident angle is set to an arbitrary fixed value, the pitch angle is set to an arbitrary fixed value, the far field is set to a single station, a reference coordinate system is set for each rotor separately, a variable representing the rotor rotation angle is added, the rotor component rotation parameters are set, the set variable representing the rotor component rotation angle is directly used as the angle parameter in the rotor component rotation parameters, the grid search parameters are set, the rotation of the quadrotor components and the change of the plane wave incident angle and pitch angle are realized, the FEKO software is run to obtain a .ffe file sequence, each .ffe file is processed, and the echo data is extracted.

[0015] The step (2) specifically refers to: the radar echo signal of the single-rotor UAV It is expressed as:

[0016] S M ( t )  = ∑ m = 0 M − 1 r ( L 2 ′ − L 1 ′ )sin c [ 2 π l ( L 2 ′ − L 1 ′ )cos β cos ( i 0 + 2 π m / M + 2 π f r t − α ) ] exp { j 2 π f c t - j 4 π l [ R + vt + L 2 ′ + L 1 ′ 2 cos β cos( i 0 + 2 π m / M + 2 π f r t - α ) ] }

[0017] in, is the total number of blades of a single-rotor drone, is the scattering coefficient, is the distance from the rotation center of the single-rotor drone to the root of the rotor blade, is the distance from the rotation center to the tip of the rotor blade of a single-rotor drone, and They represent the azimuth and elevation angles of the fuselage relative to the radar, represents the initial phase angle of the rotor blade, Indicates the number of revolutions per second of the rotor. represents the radar carrier frequency, Indicates the distance between the fuselage and the radar. is the translational speed of the drone, is the total duration of signal reception; is the wavelength of the radar signal, Represents the first leaves;

[0018] For a quadrotor drone, in the radar coordinate system middle, is the coordinate center and also the radar signal emission source; in the UAV coordinate system middle, It is the coordinate center and also the center of the fuselage; For the The center of rotation of the rotor, and The connection is set to , The length is , n=1,2,3,4; and The connection of is set to OQ, and The connection is set to ;

[0019] By OQ and get Length :

[0020]

[0021] in, for The angle with the x-axis;

[0022] therefore for The coordinates projected on the XOY plane, after projection The angle with the X axis is The azimuth relative to the radar source is set as ,but:

[0023]

[0024] In the formula, for Coordinate system with drone The angle of the axis;

[0025] set up The azimuth relative to the radar source is ,and The angle with the Z axis is ,but:

[0026]

[0027] For a quadcopter, the radar echo signal of the first rotor for:

[0028] S 1 ( t )  = ∑ k = 0 N − 1 r ( L 2 − L 1 )sin c [ 2 π l ( L 2 − L 1 )cos β 1 cos ( i 01 + 2 π k / N + 2 π f r 1 t − α 1 ) ] exp { j 2 π f c t - j 4 π l [ R op 1 + L 2 + L 1 2 cos β 1 cos( i 01 + 2 π k / N + 2 π f r 1 t - α 1 ) ] }

[0029] in, is the initial angle of the first rotor; is the total number of blades of each rotor of the quadrotor drone, The quadcopter drone leaves, is the distance from the rotation center of the quadrotor drone to the root of the rotor blade, is the distance from the rotation center of the quadrotor drone to the tip of the rotor blade, and represent the azimuth and pitch angle of the first rotor center relative to the radar, Indicates the number of revolutions per second of the first rotor;

[0030] Radar echo signal of the second rotor for:

[0031] S 2 ( t )  = ∑ k = 0 N − 1 r ( L 2 − L 1 )sin c [ 2 π l ( L 2 − L 1 )cos β 2 cos ( i 02 + 2 π k / N + 2 π f r 2 t − α 2 ) ] exp { j 2 π f c t - j 4 π l [ R op 2 + L 2 + L 1 2 cos β 2 cos( i 02 + 2 π k / N + 2 π f r 2 t - α 2 ) ] }

[0032] In the formula, and represent the azimuth and pitch angle of the second rotor center relative to the radar, is the initial angle of the second rotor, Indicates the number of revolutions per second of the second rotor;

[0033] Radar echo signal of the third rotor for:

[0034] S 3 ( t )  = ∑ k = 0 N − 1 r ( L 2 − L 1 )sin c [ 2 π l ( L 2 − L 1 )cos β 3 cos ( i 03 + 2 π k / N + 2 π f r 3 t − α 3 ) ] exp { j 2 π f c t - j 4 π l [ R op 3 + L 2 + L 1 2 cos β 3 cos( i 03 + 2 π k / N + 2 π f r 3 t - α 3 ) ] }

[0035] In the formula, and They represent the azimuth and pitch angle of the third rotor center relative to the radar, is the initial angle of the third rotor, Indicates the number of revolutions per second of the third rotor;

[0036] Radar echo signal of the fourth rotor for:

[0037] S 4 ( t )  = ∑ k = 0 N − 1 r ( L 2 − L 1 )sin c [ 2 π l ( L 2 − L 1 )cos β 4 cos ( i 01 + 2 π k / N + 2 π f r 4 t − α 4 ) ] exp { j 2 π f c t - j 4 π l [ R op 4 + L 2 + L 1 2 cos β 4 cos( i 01 + 2 π k / N + 2 π f r 4 t - α 4 ) ] }

[0038] In the formula, and They represent the azimuth and pitch angles of the fourth rotor center relative to the radar, is the initial angle of the fourth rotor, Indicates the number of revolutions per second of the fourth rotor;

[0039] The radar echo signals of the four rotors , , , The addition gives the echo signal of the quad-rotor drone.

[0040] The step (3) specifically refers to: given the original signal and window function , the definition of short-time Fourier transform is as follows:

[0041]

[0042] In the formula, represents the angular frequency, For time, express middle Pick The result of Express The signal is obtained in reverse order.

[0043] The step (5) specifically refers to: the ResNet18-SVM neural network based on the attention mechanism is composed of an input layer, a ResNet18 network, a convolutional attention module, an SVM classifier and an output layer, wherein the ResNet18 network includes a convolutional layer, four residual sub-networks and a fully connected layer, the convolutional kernel size of the convolutional layer is 3×3, the step size is 1, and the padding is 1; the residual sub-network includes two convolutional layers, a batch normalization layer and a ReLU activation function layer, the convolutional kernel size of the convolutional layer is 3×3, and the step size is 1; the convolutional attention module includes a spatial attention mechanism and a channel attention mechanism module, the convolutional attention module is added after the last residual sub-network of the ResNet18 network, and the fully connected layer of the ResNet18 network is replaced with an SVM classifier.

[0044] It can be seen from the above technical scheme that the beneficial effects of the present invention are: first, the method of using FEKO to simulate single-rotor UAV and multi-rotor UAV radar echoes and the method of using MATLAB theory to calculate multi-rotor UAV radar echoes proposed in the present invention can replace actual measurement, and obtain more accurate UAV radar echoes more conveniently and at a lower cost; second, the ResNet18 network architecture of the present invention effectively solves the gradient vanishing or gradient exploding problems that may occur in the training process of deep neural networks by introducing residual connections, so that the network can extract features at a deeper level, thereby improving the recognition accuracy; third, the present invention introduces an attention mechanism so that the network can focus on the key information in the input data, namely the micro-Doppler characteristics of the aircraft, which is crucial for target recognition in complex backgrounds; fourth, the present invention adopts support vector machine SVM as the application of the classifier, which further enhances the classification performance of the network, so that the model can still maintain a high recognition accuracy when facing diversified and complex aircraft targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of the method of the present invention;

[0046] Figure 2 It is a rotor model diagram of a single-rotor four-blade UAV of the present invention;

[0047] Figure 3 It is a rotor model diagram of a single-rotor three-blade UAV of the present invention;

[0048] Figure 4 It is a diagram of the relative positions of the quadrotor drone and the radar signal source;

[0049] Figure 5 This is the time-frequency diagram of the radar echo signal of a single-rotor four-blade UAV simulated by FEKO after short-time Fourier transform;

[0050] Figure 6 This is the time-frequency diagram of the radar echo signal of a single-rotor three-blade UAV simulated by FEKO after short-time Fourier transform;

[0051] Figure 7 This is the time-frequency diagram of the radar echo signal of the quad-rotor drone simulated by FEKO after short-time Fourier transform;

[0052] Figure 8 It is the time-frequency diagram obtained by short-time Fourier transform of the radar echo signal of a single-rotor four-blade UAV through theoretical calculation in MATLAB;

[0053] Figure 9 It is the time-frequency diagram obtained by short-time Fourier transform of the radar echo signal of a single-rotor three-blade UAV through theoretical calculation in MATLAB;

[0054] Figure 10 It is the time-frequency diagram obtained by short-time Fourier transform of the radar echo signal of the quad-rotor drone through theoretical calculation in MATLAB;

[0055] Figure 11 It is a structural diagram of the ResNet18-SVM neural network based on the attention mechanism in the present invention;

[0056] Figure 12 It is the confusion matrix of classifying the time-frequency graphs of three types of rotorcraft using the trained ResNet18-SVM neural network based on the attention mechanism. DETAILED DESCRIPTION

[0057] like Figure 1 As shown, a rotorcraft target recognition method based on micro-Doppler features and deep learning includes the following steps in sequence:

[0058] (1) Simulate the echo signals of a single-rotor four-blade UAV, a single-rotor three-blade UAV, and a four-rotor UAV respectively;

[0059] (2) The echo signals of a single-rotor four-blade UAV, a single-rotor three-blade UAV, and a four-rotor UAV are obtained through theoretical calculations;

[0060] (3) The echo signals described in steps (1) and (2) are transformed into time-frequency signals by short-time Fourier transform, and a time-frequency diagram data set representing the micro-Doppler characteristics of a single-rotor four-blade UAV, a single-rotor three-blade UAV, and a quad-rotor UAV is obtained. The time-frequency diagram of the single-rotor four-blade UAV has a micro-Doppler curve that is symmetrical up and down, the micro-Doppler curve in the time-frequency diagram of the single-rotor three-blade UAV is asymmetrical, and the micro-Doppler curve generated by each rotor in the time-frequency diagram of the quad-rotor UAV is symmetrical. Due to the phase difference of each rotor and the difference in the initial angle of the blades, the micro-Doppler curves generated by the four rotors are in an interlaced shape, which is consistent with the theoretical practice.

[0061] (4) Divide the time-frequency graph dataset into a training set and a test set in a ratio of 8:2;

[0062] (5) Construct a ResNet18-SVM neural network based on the attention mechanism;

[0063] (6) Train the ResNet18-SVM neural network based on the attention mechanism through the training set, with a batch size of 128, a training round number of 10, and a learning rate of 0.01 to obtain the trained ResNet18-SVM neural network based on the attention mechanism, and test the ResNet18-SVM neural network based on the attention mechanism through the test set;

[0064] (7) Use the trained ResNet18-SVM neural network based on the attention mechanism to classify the time-frequency graphs of the three types of rotorcraft and realize target recognition of the three types of rotorcraft.

[0065] like Figure 2 , Figure 3 As shown, the step (1) specifically refers to: establishing a single-rotor four-blade UAV model and a single-rotor three-blade UAV model and importing them into the FEKO software, placing the rotor center at the origin of the coordinate system, setting the incident wave frequency, setting the pitch angle to an arbitrary fixed value, setting the far field to a single station, setting the plane wave azimuth angle to start from a certain initial angle, uniformly sampling to the range of the initial angle plus 360°, simulating the rotation of the rotor with the change of the plane wave azimuth angle, using the grid search function to change the initial angle and pitch angle of the plane wave azimuth angle, running the FEKO software to obtain a .ffe file sequence, processing each .ffe file, and extracting echo data;

[0066] For a quadrotor drone, a quadrotor drone model is established and imported into the FEKO software, the incident wave frequency is set, the plane wave incident angle is set to an arbitrary fixed value, the pitch angle is set to an arbitrary fixed value, the far field is set to a single station, a reference coordinate system is set for each rotor separately, a variable representing the rotor rotation angle is added, the rotor component rotation parameters are set, the set variable representing the rotor component rotation angle is directly used as the angle parameter in the rotor component rotation parameters, the grid search parameters are set, the rotation of the quadrotor components and the change of the plane wave incident angle and pitch angle are realized, the FEKO software is run to obtain a .ffe file sequence, each .ffe file is processed, and the echo data is extracted.

[0067] The step (2) specifically refers to: the radar echo signal of the single-rotor UAV It is expressed as:

[0068] S M ( t )  = ∑ m = 0 M − 1 r ( L 2 ′ − L 1 ′ )sin c [ 2 π l ( L 2 ′ − L 1 ′ )cos β cos ( i 0 + 2 π m / M + 2 π f r t − α ) ] exp { j 2 π f c t - j 4 π l [ R + vt + L 2 ′ + L 1 ′ 2 cos β cos( i 0 + 2 π m / M + 2 π f r t - α ) ] }

[0069] in, is the total number of blades of a single-rotor drone, is the scattering coefficient, is the distance from the rotation center of the single-rotor drone to the root of the rotor blade, is the distance from the rotation center to the tip of the rotor blade of a single-rotor drone, and They represent the azimuth and elevation angles of the fuselage relative to the radar, represents the initial phase angle of the rotor blade, Indicates the number of revolutions per second of the rotor. represents the radar carrier frequency, Indicates the distance between the fuselage and the radar. is the translational speed of the drone, is the total duration of signal reception; is the wavelength of the radar signal, Indicates the first leaves;

[0070] like Figure 4 As shown, for a quad-rotor drone, in the radar coordinate system middle, is the coordinate center and also the radar signal emission source; in the UAV coordinate system middle, It is the coordinate center and also the center of the fuselage; For the The center of rotation of the rotor, and The connection is set to , The length is , n=1,2,3,4; and The connection of is set to OQ, and The connection is set to ;

[0071] By OQ and get Length :

[0072]

[0073] in, for The angle with the x-axis;

[0074] therefore for The coordinates projected on the XOY plane, after projection The angle with the X axis is The azimuth relative to the radar source is set as ,but:

[0075]

[0076] In the formula, for Coordinate system with drone The angle of the axis;

[0077] set up The azimuth relative to the radar source is ,and The angle with the Z axis is ,but:

[0078]

[0079] For a quadcopter, the radar echo signal of the first rotor for:

[0080] S 1 ( t )  = ∑ k = 0 N − 1 r ( L 2 − L 1 )sin c [ 2 π l ( L 2 − L 1 )cos β 1 cos ( i 01 + 2 π k / N + 2 π f r 1 t − α 1 ) ] exp { j 2 π f c t - j 4 π l [ R op 1 + L 2 + L 1 2 cos β 1 cos( i 01 + 2 π k / N + 2 π f r 1 t - α 1 ) ] }

[0081] in, is the initial angle of the first rotor; is the total number of blades of each rotor of the quadrotor drone, The quadcopter drone leaves, is the distance from the rotation center of the quadrotor drone to the root of the rotor blade, is the distance from the rotation center of the quadrotor drone to the tip of the rotor blade, and represent the azimuth and pitch angle of the first rotor center relative to the radar, Indicates the number of revolutions per second of the first rotor;

[0082] Radar echo signal of the second rotor for:

[0083] S 2 ( t )  = ∑ k = 0 N − 1 r ( L 2 − L 1 )sin c [ 2 π l ( L 2 − L 1 )cos β 2 cos ( i 02 + 2 π k / N + 2 π f r 2 t − α 2 ) ] exp { j 2 π f c t - j 4 π l [ R op 2 + L 2 + L 1 2 cos β 2 cos( i 02 + 2 π k / N + 2 π f r 2 t - α 2 ) ] }

[0084] In the formula, and represent the azimuth and pitch angle of the second rotor center relative to the radar, is the initial angle of the second rotor, Indicates the number of revolutions per second of the second rotor;

[0085] Radar echo signal of the third rotor for:

[0086] S 3 ( t )  = ∑ k = 0 N − 1 r ( L 2 − L 1 )sin c [ 2 π l ( L 2 − L 1 )cos β 3 cos ( i 03 + 2 π k / N + 2 π f r 3 t − α 3 ) ] exp { j 2 π f c t - j 4 π l [ R op 3 + L 2 + L 1 2 cos β 3 cos( i 03 + 2 π k / N + 2 π f r 3 t - α 3 ) ] }

[0087] In the formula, and They represent the azimuth and pitch angle of the third rotor center relative to the radar, is the initial angle of the third rotor, Indicates the number of revolutions per second of the third rotor;

[0088] Radar echo signal of the fourth rotor for:

[0089] S 4 ( t )  = ∑ k = 0 N − 1 r ( L 2 − L 1 )sin c [ 2 π l ( L 2 − L 1 )cos β 4 cos ( i 01 + 2 π k / N + 2 π f r 4 t − α 4 ) ] exp { j 2 π f c t - j 4 π l [ R op 4 + L 2 + L 1 2 cos β 4 cos( i 01 + 2 π k / N + 2 π f r 4 t - α 4 ) ] }

[0090] In the formula, and They represent the azimuth and pitch angles of the fourth rotor center relative to the radar, is the initial angle of the fourth rotor, Indicates the number of revolutions per second of the fourth rotor;

[0091] The radar echo signals of the four rotors , , , The addition results in the echo signal of the quad-rotor drone. This theoretical calculation method of radar echo takes into account the different distances, azimuths, and pitch angles of each rotor relative to the radar emission source, and can more accurately calculate the theoretical radar echo of a multi-rotor drone.

[0092] The step (3) specifically refers to: given the original signal and window function , the definition of short-time Fourier transform is as follows:

[0093]

[0094] In the formula, represents the angular frequency, For time, express middle Pick The result of Express The signal is obtained in reverse order.

[0095] The step (4) specifically refers to: randomly shuffling 1332 time-frequency images of the single-rotor four-blade UAV, 1242 time-frequency images of the single-rotor three-blade UAV, and 2736 time-frequency images of the quad-rotor UAV, and dividing them into a training set and a test set in a ratio of 8:2.

[0096] The step (5) specifically refers to: Figure 11 As shown in the figure, the ResNet18-SVM neural network based on the attention mechanism consists of an input layer, a ResNet18 network, a convolutional attention module, an SVM classifier and an output layer, wherein the ResNet18 network includes a convolutional layer, four residual sub-networks and a fully connected layer, the convolution kernel size of the convolution layer is 3×3, the step size is 1, and the padding is 1; the residual sub-network includes two convolutional layers, a batch normalization layer and a ReLU activation function layer, the convolution kernel size of the convolution layer is 3×3, and the step size is 1; the convolutional attention module includes a spatial attention mechanism and a channel attention mechanism module, and adding the convolutional attention module after the last residual sub-network of the ResNet18 network helps the network focus on the most informative features, better understand the characteristics of the time-frequency graph, and improve the classification performance; the fully connected layer of the ResNet18 network is replaced with an SVM classifier to improve the generalization ability of the model and improve boundary decision-making.

[0097] like Figure 5 , Figure 8 As shown, since the blades of a single-rotor four-blade UAV are symmetrical in pairs about the center of the rotor, the micro-Doppler time-frequency diagram of the single-rotor four-blade UAV is a sine curve that is symmetrical up and down, and two "time-frequency flickers" will occur in each cycle of the sine curve.

[0098] like Figure 6 , Figure 9 As shown in the figure, the micro-Doppler time-frequency diagram of the single-rotor three-blade UAV is asymmetric up and down, and two "time-frequency flicker" phenomena will occur in each cycle of the sine curve.

[0099] like Figure 7 , Figure 10 As shown in the figure, since each rotor of the quadcopter has two blades, the time-frequency diagram is a symmetrical sine curve. In addition, since the position of each rotor center relative to the radar signal emission source is different, and the initial phase of each rotor is different, the micro-Doppler curve generated by each rotor will produce a phase difference in the time-frequency diagram.

[0100] like Figure 12 As shown in the figure, single-rotor three-blade drones and four-rotor drones were all accurately classified, and 1% of single-rotor four-blade drones were misidentified as single-rotor three-blade drones, with good overall recognition performance. It can be seen that in the recognition of the three types of targets, only the single-rotor four-blade drone had an error of 0.01, and the final recognition rate was 99.72%.

[0101] In summary, the method of simulating single-rotor UAV and multi-rotor UAV radar echoes using FEKO and the method of calculating multi-rotor UAV radar echoes using MATLAB theory proposed in the present invention can replace actual measurement and obtain more accurate UAV radar echoes more conveniently at a lower cost; the ResNet18-SVM neural network based on the attention mechanism of the present invention has good performance in target recognition based on aircraft micro-Doppler characteristics.

Claims

1. A rotorcraft target recognition method based on micro-Doppler features and deep learning, characterized in that: The method comprises the following steps in order: (1) Simulate the echo signals of a single-rotor four-blade UAV, a single-rotor three-blade UAV, and a four-rotor UAV respectively; (2) The echo signals of a single-rotor four-blade UAV, a single-rotor three-blade UAV, and a four-rotor UAV are obtained through theoretical calculations; (3) performing time-frequency transformation on the echo signals described in steps (1) and (2) by short-time Fourier transform, and obtaining a time-frequency graph data set representing micro-Doppler characteristics of a single-rotor four-blade UAV, a single-rotor three-blade UAV, and a quad-rotor UAV; (4) Divide the time-frequency graph dataset into a training set and a test set in a ratio of 8:2; (5) Construct a ResNet18-SVM neural network based on the attention mechanism; (6) Train the ResNet18-SVM neural network based on the attention mechanism through the training set to obtain the trained ResNet18-SVM neural network based on the attention mechanism, and test the ResNet18-SVM neural network based on the attention mechanism through the test set; (7) Use the trained ResNet18-SVM neural network based on the attention mechanism to classify the time-frequency graphs of the three types of rotorcraft and realize target recognition of the three types of rotorcraft; The step (2) specifically refers to: the radar echo signal of the single-rotor UAV It is expressed as: ; in, is the total number of blades of a single-rotor drone, is the scattering coefficient, is the distance from the rotation center of the single-rotor drone to the root of the rotor blade, is the distance from the rotation center to the tip of the rotor blade of a single-rotor drone, and They represent the azimuth and elevation angles of the fuselage relative to the radar, represents the initial phase angle of the rotor blade, Indicates the number of revolutions per second of the rotor. represents the radar carrier frequency, Indicates the distance between the fuselage and the radar. is the translational speed of the drone, is the total duration of signal reception; is the wavelength of the radar signal, Indicates the first leaves; For a quadrotor drone, in the radar coordinate system middle, is the coordinate center and also the radar signal emission source; in the UAV coordinate system middle, It is the coordinate center and also the center of the fuselage; For the The center of rotation of the rotor, and The connection is set to , The length is , n=1,2,3,4; and The connection of is set to OQ, and The connection is set to ; By OQ and get Length : ; in, for The angle with the x-axis; therefore for The coordinates projected on the XOY plane, after projection The angle with the X axis is The azimuth relative to the radar source is set as ,but: ; In the formula, for Coordinate system with drone The angle of the axes; set up The azimuth relative to the radar source is ,and The angle with the Z axis is ,but: ; For a quadcopter, the radar echo signal of the first rotor for: ; in, is the initial angle of the first rotor; is the total number of blades of each rotor of the quadrotor drone, The quadcopter drone leaves, is the distance from the rotation center of the quadrotor drone to the root of the rotor blade, is the distance from the rotation center of the quadrotor drone to the tip of the rotor blade, and They represent the azimuth and pitch angle of the first rotor center relative to the radar, Indicates the number of revolutions per second of the first rotor; Radar echo signal of the second rotor for: ; In the formula, and represent the azimuth and pitch angle of the second rotor center relative to the radar, is the initial angle of the second rotor, Indicates the number of revolutions per second of the second rotor; Radar echo signal of the third rotor for: ; In the formula, and They represent the azimuth and pitch angle of the third rotor center relative to the radar, is the initial angle of the third rotor, Indicates the number of revolutions per second of the third rotor; Radar echo signal of the fourth rotor for: ; In the formula, and They represent the azimuth and pitch angles of the fourth rotor center relative to the radar, is the initial angle of the fourth rotor, Indicates the number of revolutions per second of the fourth rotor; The radar echo signals of the four rotors , , , The echo signal of the quad-rotor drone is obtained by adding them together. This theoretical calculation method of radar echo takes into account that the distance, azimuth and pitch angle of each rotor relative to the radar emission source are different, and the theoretical radar echo of the multi-rotor drone is calculated more accurately. The step (5) specifically refers to: the ResNet18-SVM neural network based on the attention mechanism is composed of an input layer, a ResNet18 network, a convolutional attention module, an SVM classifier and an output layer, wherein the ResNet18 network includes a convolutional layer, four residual sub-networks and a fully connected layer, the convolutional kernel size of the convolutional layer is 3×3, the step size is 1, and the padding is 1; the residual sub-network includes two convolutional layers, a batch normalization layer and a ReLU activation function layer, the convolutional kernel size of the convolutional layer is 3×3, and the step size is 1; the convolutional attention module includes a spatial attention mechanism and a channel attention mechanism module, and adding the convolutional attention module after the last residual sub-network of the ResNet18 network helps the network focus on the most informative features, better understands the features of the time-frequency graph, and improves the classification performance; the fully connected layer of the ResNet18 network is replaced with an SVM classifier to improve the generalization ability of the model and improve boundary decision-making.

2. The method for rotorcraft target recognition based on micro-Doppler features and deep learning according to claim 1, characterized in that: The step (1) specifically includes: establishing a single-rotor four-blade UAV model and a single-rotor three-blade UAV model and importing them into FEKO software, placing the rotor center at the origin of the coordinate system, setting the incident wave frequency, setting the pitch angle to an arbitrary fixed value, setting the far field to a single station, setting the plane wave azimuth angle to start from a certain initial angle, uniformly sampling to a range of the initial angle plus 360°, simulating the rotation of the rotor with the change of the plane wave azimuth angle, using the grid search function to change the initial angle and pitch angle of the plane wave azimuth angle, running the FEKO software to obtain a .ffe file sequence, processing each .ffe file, and extracting echo data; For a quadrotor drone, a quadrotor drone model is established and imported into the FEKO software, the incident wave frequency is set, the plane wave incident angle is set to an arbitrary fixed value, the pitch angle is set to an arbitrary fixed value, the far field is set to a single station, a reference coordinate system is set for each rotor separately, a variable representing the rotor rotation angle is added, the rotor component rotation parameters are set, the set variable representing the rotor component rotation angle is directly used as the angle parameter in the rotor component rotation parameters, the grid search parameters are set, the rotation of the quadrotor components and the change of the plane wave incident angle and pitch angle are realized, the FEKO software is run to obtain a .ffe file sequence, each .ffe file is processed, and the echo data is extracted.

3. The method for rotorcraft target recognition based on micro-Doppler features and deep learning according to claim 1, characterized in that: The step (3) specifically refers to: given the original signal and window function , the definition of short-time Fourier transform is as follows: ; In the formula, represents the angular frequency, For time, express middle Pick The result of Express The signal is obtained in reverse order.

Citation Information

Cited By

  • UAV reconnaissance and countermeasure method based on deep learning

    CN119814219B