CN-121982370-A - Unmanned aerial vehicle cluster type recognition method and device based on deep learning

CN121982370ACN 121982370 ACN121982370 ACN 121982370ACN-121982370-A

Abstract

The invention discloses an unmanned aerial vehicle cluster type recognition method based on deep learning, which comprises the steps of establishing a cluster radar echo signal model according to a micro Doppler effect and an electromagnetic scattering principle, characterizing echo signals as superposition of echo waves of unmanned aerial vehicles in a cluster by the model, performing time-frequency analysis on the echo signals, generating a micro Doppler time-frequency chart and mapping the micro Doppler time-frequency chart into characteristic images, constructing a data set after labeling types, constructing ResNet network models, relieving gradient disappearance by a residual error module, extracting low-dimensional visual characteristics and high-level semantic characteristics by multi-scale convolution, realizing classification by characteristic fusion and a full-connection layer, and inputting the characteristic images generated by real-time echo signals into the trained models to complete cluster type recognition. The method can effectively capture the micro Doppler characteristic of the cluster, has high recognition precision and strong generalization, and provides reliable technical support for unmanned aerial vehicle cluster recognition.

Inventors

CHEN ZIHAN
DANG XIAOYU
YU XIANGBIN
LI SAI
Cheng Beien
SONG CHAOHUI

Assignees

南京航空航天大学
杭州电子科技大学

Dates

Publication Date: 20260505
Application Date: 20251225

Claims (6)

1. The unmanned aerial vehicle cluster type recognition method based on deep learning is characterized by comprising the following steps of: S1, a cluster radar echo signal model is established according to the micro Doppler effect and the electromagnetic scattering correlation principle, wherein radar echo signals of the unmanned aerial vehicle cluster received by a radar receiving end are superposition of radar echo waves of each unmanned aerial vehicle in the cluster; S2, performing time-frequency analysis on radar echo signals of the unmanned aerial vehicle cluster to generate a time-frequency data matrix and constructing a micro Doppler time-frequency diagram, mapping signal energy values in the micro Doppler time-frequency diagram into color levels in an optical image, enabling the color of each pixel to represent energy intensity of different frequency components in a micro Doppler component, and constructing a cluster characteristic data sample after marking unmanned aerial vehicle cluster types on the generated unmanned aerial vehicle cluster micro Doppler characteristic image; S3, constructing a ResNet network model based on micro Doppler features, performing end-to-end training on the ResNet network model by using cluster feature data samples, wherein the ResNet network model relies on a residual error learning module to solve the gradient disappearance problem of a depth network, firstly extracting multi-scale space features of an input unmanned aerial vehicle cluster micro Doppler feature image through a stacked convolution layer, a batch normalization layer and a ReLU activation function, capturing low-dimensional visual features including vertical stripe edges and pixel intensity distribution in the image by a shallow convolution layer, fusing and abstracting high-level semantic features including the shape, the width and the energy distribution of micro Doppler frequency bands by a deep residual error block, performing dimension alignment and weighted fusion on the fused high-dimensional features by a feature fusion module, performing dimension reduction and classification mapping on the feature fusion by using a full-connection layer, finally outputting identification probability of unmanned aerial vehicle cluster types based on classification results, and optimizing network parameters by back propagation iteration to improve identification accuracy; s4, importing the unmanned aerial vehicle cluster micro Doppler characteristic images corresponding to the radar echo signals of the unmanned aerial vehicle clusters acquired in real time into a ResNet network model with training completed, and identifying the unmanned aerial vehicle cluster target types.
2. The deep learning-based unmanned aerial vehicle cluster type recognition method according to claim 1, wherein step S1 further comprises: using constant radial velocities respectively Sum angular velocity Modeling movements of unmanned aerial vehicle bodies and micro-movements of blades for radial velocity Distance from scattering point P on rotor blade of moving unmanned plane to radar at time t The method comprises the following steps: ; The radar echo signal returned from point scatterer P at time t is: ; Wherein, the Representing a phase function of the point scatterer; Representing a carrier frequency; Representing the distance from the scattering point P to the rotation center on the rotor, the equivalent instantaneous micro Doppler frequency of the point scatterer P is: ; The baseband signal returned by the point scatterer P is a single component signal: ; Integrating the length L of the whole blade to obtain a radar echo signal of a single blade, wherein the radar echo signal is as follows: ; Wherein, the , Representing a non-normalized Sinc function; For a single rotor unmanned aerial vehicle having N blades, the initial rotation angle of each blade is different, then the initial rotation angle of the kth blade is expressed as: ; The radar returns of the single rotor unmanned aerial vehicle are expressed as: ; Wherein, the , Indicating the initial rotation angle of the rotor, And Representing the azimuth and elevation of the drone relative to the radar, The initial distance between the unmanned aerial vehicle target and the radar is set; The expression for deriving the radar echo of the multi-rotor unmanned aerial vehicle by superposing the radar echo of the single rotor blade is as follows: ; The phase function is: ; Wherein, the The number of the rotor wings of the unmanned aerial vehicle is represented, For the number of blades on each rotor, Indicating the length of the blade and, And (3) with Is the first The initial distance and the initial rotation angle between the individual rotors and the radar, For the rotation speed of the r-th rotor, And Represent the first Azimuth and elevation angles of the individual rotors relative to the radar; and obtaining the equivalent instantaneous micro Doppler frequency of the kth blade scattering point P on the kth rotor wing by the time derivative of the phase function, wherein the equivalent instantaneous micro Doppler frequency is as follows: ; when M unmanned aerial vehicles exist in the unmanned aerial vehicle cluster, radar echo signal modeling of the unmanned aerial vehicle cluster is as follows: ; The phase function is: ; Wherein the method comprises the steps of , 、 Respectively representing the rotation angular velocity and the initial rotation angle of the (r) th rotor wing of the (i) th unmanned aerial vehicle in the cluster; 、 A is a real number amplitude scaling factor, and is related to radar transmitting power, radar sectional area of a target and antenna gain; the number of the rotor wings of the single unmanned aerial vehicle in the cluster is represented, Representing the number of blades on the rotor of the drone within the cluster, Representing an initial distance between an (r) th rotor of the (i) th unmanned aerial vehicle and the radar; and obtaining the equivalent instantaneous micro Doppler frequency of the kth blade scattering point P on the ith rotor wing of the ith unmanned aerial vehicle by the time derivative of the phase function, wherein the equivalent instantaneous micro Doppler frequency is as follows: ; considering noise influence, defining radar echo signals of the unmanned aerial vehicle cluster as: ; Wherein, the Representing the radar echo of the ith drone, Mean value is zero and variance is Additive white gaussian noise of (c).
3. The deep learning-based unmanned aerial vehicle cluster type recognition method according to claim 1, wherein in step S2, the unmanned aerial vehicles in the unmanned aerial vehicle cluster include single-rotor unmanned aerial vehicles, four-rotor unmanned aerial vehicles, six-rotor unmanned aerial vehicles and eight-rotor unmanned aerial vehicles.
4. The unmanned aerial vehicle cluster type recognition method based on deep learning according to claim 1, wherein in step S2, a short-time fourier transform is adopted to perform time-frequency analysis on original echo signals, and radar echo signals of unmanned aerial vehicle clusters are obtained The time-frequency analysis result of (2) is expressed as: ; Wherein, the As a window function.
5. Unmanned aerial vehicle cluster type recognition device based on deep learning, characterized in that the device includes: The cluster radar echo signal model construction module is used for building a cluster radar echo signal model according to the micro Doppler effect and the electromagnetic scattering correlation principle, wherein radar echo signals of the unmanned aerial vehicle cluster received by a radar receiving end are superposition of radar echo waves of each unmanned aerial vehicle in the cluster; The system comprises a micro Doppler characteristic image generation module, a micro Doppler characteristic image generation module and a micro Doppler characteristic image generation module, wherein the micro Doppler characteristic image generation module is used for carrying out time-frequency analysis on radar echo signals of an unmanned aerial vehicle cluster to generate a time-frequency data matrix and construct a micro Doppler time-frequency image; the training sample construction module is used for marking unmanned aerial vehicle cluster types on the generated unmanned aerial vehicle cluster micro Doppler characteristic images and then constructing cluster characteristic data samples; The system comprises a ResNet network model training module, a ResNet network model, a feature fusion module, a full-connection layer, a classification mapping module and a recognition module, wherein the ResNet network model training module is used for constructing a ResNet network model based on micro Doppler features, performing end-to-end training on the ResNet network model by utilizing cluster feature data samples, solving the gradient disappearance problem of a depth network by relying on the residual error learning module, firstly extracting multi-scale spatial features of an input unmanned aerial vehicle cluster micro Doppler feature image by using a stacked convolution layer, a batch normalization layer and a ReLU activation function, capturing low-dimensional visual features including vertical stripe edges and pixel intensity distribution in the image by using a shallow convolution layer, fusing and abstracting high-level semantic features including the shape, the width and the energy distribution of a micro Doppler frequency band by using a deep residual error block, performing dimension alignment and weighting fusion on the features of different levels by using the feature fusion module, finally outputting recognition probability of unmanned aerial vehicle cluster types by using a full-connection layer and optimizing network parameters by reverse propagation iteration so as to improve recognition accuracy; And the ResNet network model processes unmanned aerial vehicle cluster micro Doppler characteristic images corresponding to the real-time acquired unmanned aerial vehicle cluster radar echo signals input by the micro Doppler characteristic image generation module, and identifies unmanned aerial vehicle cluster target types.
6. A computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when executed by at least one processor, causes the at least one processor to implement the unmanned aerial vehicle cluster micro doppler feature identification method as claimed in any one of claims 1-4.

Description

Unmanned aerial vehicle cluster type recognition method and device based on deep learning Technical Field The invention relates to the technical field of radar target identification, in particular to an unmanned aerial vehicle cluster type identification method and device based on deep learning. Background In recent years, along with the rapid development of unmanned aerial vehicle technology, unmanned aerial vehicles are widely applied in civil and military fields by virtue of the characteristics of relatively low cost, flexibility, strong concealment, capability of executing tasks in complex environments and the like. However, single-frame unmanned aerial vehicles still have certain limitations in terms of task load, coverage, viability, coordination with complex tasks, and the like. Aiming at the requirements of higher-intensity and more complex tasks, the adoption of a clustering execution mode of multi-unmanned aerial vehicle cooperation has become one of important development directions. Compared with the rapid evolution of the related technology of the unmanned aerial vehicle cluster, the detection, identification and countering technology for the unmanned aerial vehicle cluster is still relatively lagged. In order to realize effective defense and management of the unknown unmanned aerial vehicle clusters, the cluster types and the organization forms of the unknown unmanned aerial vehicle clusters are generally required to be identified firstly, and corresponding treatment strategies are adopted by combining the motion and the cooperative characteristics of different clusters on the basis. Therefore, the development of the recognition technology research oriented to different unmanned aerial vehicle clusters is of great significance. The existing detection and classification means for the small unmanned aerial vehicle mainly comprise radar detection, optical imaging, acoustic detection and the like. The radar has all-weather and all-day working capacity and a certain long-distance detection advantage, and can acquire target echo information under complex illumination and meteorological conditions, so that the radar has important application value in detection and identification of a small unmanned aerial vehicle. In the radar observation process, the periodic rotation of the unmanned aerial vehicle rotor blade can generate periodic frequency modulation on the backward scattering echo, so that a micro Doppler effect is formed. The micro Doppler effect is closely related to the rotor rotation speed, the length of the blades, the number of the blades and other structures and dynamic parameters, and the motion characteristics of the micro component of the unmanned aerial vehicle can be represented to a certain extent. By performing time-frequency analysis on the differential Doppler component in the echo and extracting relevant characteristics (such as instantaneous frequency track, energy distribution and the like), effective information support can be provided for detection and identification of the unmanned aerial vehicle target. Aiming at the unmanned aerial vehicle identification problem based on micro Doppler characteristics, the prior art has conducted more research work. The method adopts the means of empirical mode decomposition (EMPIRICAL MODE DECOMPOSITION, EMD), singular value decomposition, principal component analysis or independent component analysis and the like to decompose and dimension down received waves, extracts statistically irrelevant or relatively independent characteristic quantities for classification, also researches and combines a sequence modeling method, such as Long Short-Term Memory (LSTM), to realize the detection and classification of unmanned aerial vehicle targets by utilizing the modeling capability of the Long Short-Term Memory (LSTM) to the time sequence, and obtains better recognition effect under certain experimental conditions. The method generally relies on analysis of micro Doppler information and characteristic engineering design, and although the method can obtain certain precision, the method has the problems of limited expandability, strong dependence on domain knowledge and the like, and meanwhile, the characteristic extraction and screening process can bring additional calculation overhead and processing delay, so that the method is unfavorable for engineering deployment of a real-time identification system. Furthermore, in the unmanned aerial vehicle cluster scene, the increase of the number of targets and the cooperative behavior can lead to the superposition of multi-target echoes and the coupling of a time-frequency structure, so that the signal characterization is more complex, and if the identification is still carried out by taking the thought of single-target inching feature extraction as the main idea, the stable characterization of the organization form and type difference of the cluster layer is difficult. For example, researchers have propose