CN-121547107-B - Unmanned aerial vehicle radio frequency signal identification method, device, equipment and medium based on improved capsule network model

CN121547107BCN 121547107 BCN121547107 BCN 121547107BCN-121547107-B

Abstract

The application discloses an unmanned aerial vehicle radio frequency signal identification method, device, equipment and medium based on an improved capsule network model, and relates to the technical field of unmanned aerial vehicle identification, wherein the method comprises the steps of generating a radio frequency time-frequency diagram through collecting an unmanned aerial vehicle radio frequency original signal and performing short-time Fourier transform and logarithmic transform; and inputting the unmanned aerial vehicle identification result into an improved capsule network model, and outputting the unmanned aerial vehicle identification result through multi-scale feature extraction, spatial relationship capturing and classification processing. Through the combination of the improved Inception structure and the capsule network, the adaptability of the model to the complex electromagnetic environment is enhanced, the accuracy and stability of unmanned aerial vehicle radio frequency signal identification are improved, the input of a large-size time-frequency diagram is adapted, the multi-scale characteristics are comprehensively extracted, and the spatial relationship is reserved.

Inventors

QIN YEMEI
GAO LIPING
Tang Yuexuan
NI WENZHI
XU XUNYAO
YANG LINGLING

Assignees

湖南工商大学

Dates

Publication Date: 20260508
Application Date: 20260121

Claims (9)

1. An unmanned aerial vehicle radio frequency signal identification method based on an improved capsule network model, which is characterized by comprising the following steps: Collecting unmanned aerial vehicle radio frequency original signals with different distances, wherein the unmanned aerial vehicle radio frequency original signals refer to electromagnetic signals which are transmitted by unmanned aerial vehicles through communication links in the flight process and are not subjected to any processing, the electromagnetic signals comprise communication instructions and state data of the unmanned aerial vehicles, the unmanned aerial vehicle radio frequency original signals are core data bases for identifying the types of the unmanned aerial vehicles, and unique frequencies and amplitudes can be displayed according to the types and the flight states of the unmanned aerial vehicles; Performing short-time Fourier transform and logarithmic transformation on the unmanned aerial vehicle radio frequency original signal to obtain a radio frequency time-frequency diagram; Inputting the radio frequency time-frequency diagram into an improved capsule network model for feature extraction and classification to obtain an unmanned aerial vehicle identification result, wherein the improved capsule network model comprises a 7×7 convolution layer, an improved Inception feature extraction module, a primary capsule layer and a digital capsule layer, and the improved Inception feature extraction module comprises three improved Inception structures; The step of inputting the radio frequency time-frequency diagram into an improved capsule network model for feature extraction and classification to obtain an unmanned aerial vehicle identification result comprises the following steps: Inputting the radio frequency time-frequency diagram into a 7 multiplied by 7 convolution layer of an improved capsule network model to perform preliminary feature extraction to obtain an initial feature diagram; processing the initial feature map through an improved Inception structure of an improved Inception feature extraction module to obtain a fusion feature map; The fusion feature map is input into a primary capsule layer to be converted into a vector capsule set; Processing the vector capsule set through a dynamic routing mechanism to obtain feature space information; And inputting the characteristic space information into a digital capsule layer for classification processing, and outputting an unmanned aerial vehicle recognition result.
2. The method of claim 1, wherein the step of processing the initial feature map through the modified Inception structure of the modified Inception feature extraction module to obtain a fused feature map comprises: Inputting the initial feature map into a first parallel path, a second parallel path, a third parallel path and a fourth parallel path of a Inception structure, wherein the first parallel path is provided with a1×1 convolution layer, the second parallel path is provided with a1×1 convolution layer and a 3×3 convolution layer in sequence, the third parallel path is provided with a1×1 convolution layer and two continuous 3×3 convolution layers in sequence, and the fourth parallel path is provided with a 3×3 pooling layer and a1×1 convolution layer in sequence; performing dimension reduction and feature recombination on the initial feature map through a1 multiplied by 1 convolution layer of the first parallel path to obtain a first dimension feature; Channel compression is carried out on the initial feature map through a 1X 1 convolution layer of the second parallel path, and then local detail association features are captured through a 3X 3 convolution layer, so that second dimension features are obtained; Feature screening is carried out on the initial feature map through a 1X 1 convolution layer of the third parallel path, and then global association features are captured through two continuous 3X 3 convolution layers, so that third dimensional features are obtained; the initial feature map is subjected to downsampling compression through a 3 multiplied by 3 pooling layer of the fourth parallel path, and then channel optimization is performed through a1 multiplied by 1 convolution layer, so that fourth-dimension features are obtained; respectively carrying out batch normalization processing on the first dimension characteristic, the second dimension characteristic, the third dimension characteristic and the fourth dimension characteristic to obtain a normalized first dimension characteristic, a normalized second dimension characteristic, a normalized third dimension characteristic and a normalized fourth dimension characteristic; And fusing the normalized first dimension characteristic, the normalized second dimension characteristic, the normalized third dimension characteristic and the normalized fourth dimension characteristic by adopting channel splicing operation to obtain a fused characteristic diagram.
3. The method of claim 1, wherein the step of converting the fused feature map input primary capsule layer into a set of vector capsules comprises: The method comprises the steps of obtaining the number of capsules of a primary capsule layer and the number of capsule types corresponding to each capsule, wherein the number of capsules is determined according to the space dimension calculation result of convolution operation, the capsule dimensions are fixed values, the space dimension calculation result of the convolution operation refers to the product of the height and the width of an output feature map after the input feature map is processed by the convolution layer, and the number of capsule types is the number of capsule types which exist in parallel at each space position in the primary capsule layer; Performing feature mapping on the multi-scale features of the fusion feature map by using a convolution layer with a preset size to obtain a multi-channel feature map; Performing dimension splitting on the multi-channel feature map based on the capsule dimension of the primary capsule layer to obtain feature blocks corresponding to the number of capsules; performing vector conversion processing on each characteristic block to obtain initial vectors corresponding to the number of capsules; Carrying out normalized compression on each initial vector through a nonlinear activation function to obtain a normalized vector capsule; and integrating all the normalized vector capsules to obtain a vector capsule set.
4. The method of claim 1, wherein the step of processing the set of vector capsules by a dynamic routing mechanism to obtain feature space information comprises: Acquiring the number of target capsules and vector dimensions preset by a digital capsule layer, and initializing an association weight matrix of each vector capsule in a target capsule and vector capsule set; Calculating initial association weights of each vector capsule and each target capsule in the vector capsule set based on the association weight matrix; Performing softmax normalization operation on all initial association weights corresponding to the same target capsule to obtain a coupling coefficient; Carrying out weighted summation on the corresponding vector capsules in the vector capsule set according to the coupling coefficient to obtain an original input vector of each target capsule; Nonlinear transformation is carried out on the original input vector through a preset compression function, and an intermediate output vector of the target capsule is obtained; calculating the vector similarity of the intermediate output vector and the corresponding vector capsule, and iteratively updating the coupling coefficient based on the similarity to obtain a coupling coefficient variation until the preset iteration times are reached or the coupling coefficient variation is smaller than a preset threshold value, so as to obtain an optimized target capsule output vector; and integrating all the optimized target capsule output vectors to obtain feature space information containing the space association relation among the features.
5. The method of claim 1, wherein the step of inputting the feature space information into a digital capsule layer for classification processing and outputting the unmanned aerial vehicle recognition result comprises the steps of: Acquiring preset parameters of a digital capsule layer, wherein the preset parameters comprise the number of digital capsules and vector dimensions of each digital capsule, and the number of the digital capsules corresponds to the number of unmanned aerial vehicle categories set by an identification task; Performing feature distribution on the feature space information according to the vector dimension to obtain category association features corresponding to each digital capsule; Performing linear transformation on each category association feature through a feature mapping matrix of the digital capsule layer to obtain a category feature vector corresponding to each digital capsule; Calculating the modular length of each category characteristic vector, and obtaining a predicted probability value of the corresponding unmanned aerial vehicle category, wherein the modular length value and the predicted probability value are positively correlated; integrating the prediction probability values of all unmanned aerial vehicle categories to generate a category prediction probability set; and screening the category with the maximum prediction probability value from the category prediction probability set to obtain an unmanned aerial vehicle recognition result.
6. The method of claim 1, wherein the step of performing short-time fourier transform and logarithmic transform on the rf raw signal of the drone to obtain an rf time-frequency diagram comprises: slicing the radio frequency original signal of the unmanned aerial vehicle, and dividing the radio frequency original signal into a plurality of continuous signal segments; Applying a Hamming window to each signal segment to obtain windowed signal segments; Performing local Fourier transform on each signal segment to obtain a corresponding local Fourier transform result; Repeating windowing and local Fourier transform operations on unprocessed signal areas in all the signal fragments according to a preset step length sliding Hamming window to obtain a full-scale local Fourier transform result; Splicing all the full local Fourier transform results according to time sequence to form a two-dimensional time-frequency matrix and obtain the joint distribution characteristics of signals; And carrying out logarithmic transformation on the joint distribution characteristics, and mapping the numerical value of the two-dimensional time-frequency matrix to a preset dynamic range to obtain a radio frequency time-frequency diagram.
7. Unmanned aerial vehicle radio frequency signal recognition device based on improve capsule network model, characterized in that, the device includes: The acquisition module is used for acquiring unmanned aerial vehicle radio frequency original signals with different distances, wherein the unmanned aerial vehicle radio frequency original signals refer to electromagnetic signals which are transmitted by the unmanned aerial vehicle through a communication link in the flight process and are not subjected to any processing, the electromagnetic signals comprise communication instructions and state data of the unmanned aerial vehicle, the unmanned aerial vehicle radio frequency original signals are core data bases for identifying the type of the unmanned aerial vehicle, and unique frequency and amplitude can be displayed according to the type and the flight state of the unmanned aerial vehicle; the processing module is used for carrying out short-time Fourier transform and logarithmic transformation on the unmanned aerial vehicle radio frequency original signal to obtain a radio frequency time-frequency diagram; The method comprises the steps of inputting a radio frequency time-frequency diagram into an improved capsule network model to conduct feature extraction and classification to obtain an unmanned aerial vehicle identification result, wherein the improved capsule network model comprises a 7×7 convolution layer, an improved Inception feature extraction module, a primary capsule layer and a digital capsule layer, the improved Inception feature extraction module comprises three improved Inception structures, the method is further used for inputting the radio frequency time-frequency diagram into the 7×7 convolution layer of the improved capsule network model to conduct primary feature extraction to obtain an initial feature diagram, processing the initial feature diagram through the improved Inception structure of the improved Inception feature extraction module to obtain a fusion feature diagram, inputting the fusion feature diagram into the primary capsule layer to be converted into a vector capsule set, processing the vector capsule set through a dynamic routing mechanism to obtain feature space information, inputting the feature space information into the digital capsule layer to conduct classification processing to obtain the unmanned aerial vehicle identification result.
8. An improved capsule network model based unmanned aerial vehicle radio frequency signal identification device, characterized in that the device comprises a memory, a processor and an improved capsule network model based unmanned aerial vehicle radio frequency signal identification program stored on the memory and running on the processor, the improved capsule network model based unmanned aerial vehicle radio frequency signal identification program being configured to implement the steps of the improved capsule network model based unmanned aerial vehicle radio frequency signal identification method of any of claims 1-6.
9. A storage medium, wherein a radio frequency signal identification program of an unmanned aerial vehicle based on an improved capsule network model is stored on the storage medium, and the radio frequency signal identification program of an unmanned aerial vehicle based on an improved capsule network model realizes the steps of the radio frequency signal identification method of an unmanned aerial vehicle based on an improved capsule network model according to any one of claims 1 to 6 when being executed by a processor.

Description

Unmanned aerial vehicle radio frequency signal identification method, device, equipment and medium based on improved capsule network model Technical Field The invention relates to the technical field of unmanned aerial vehicle identification, in particular to an unmanned aerial vehicle radio frequency signal identification method, device, equipment and medium based on an improved capsule network model. Background At present, the existing method for unmanned aerial vehicle radio frequency signal identification mainly focuses on two major directions of feature selection and deep learning model improvement. In the aspect of feature processing, a radio frequency original signal is often converted into a time-frequency diagram or a frequency-frequency diagram through methods such as short-time Fourier transform, discrete Fourier transform and wavelet transform so as to keep the time domain and frequency domain features of the signal, and in the aspect of model application in the field of unmanned plane radio frequency signal identification, the existing scheme comprises a multi-size convolutional neural network, a basic capsule network, a capsule network based on a traditional Inception structure and the like. The multi-dimensional CNN extracts multi-scale features through multi-dimensional convolution kernel stacking, a basic capsule network reserves a feature space relation by means of a dynamic routing mechanism, and the capsule network based on the traditional Inception structure combines multi-scale parallel feature extraction and capsule network advantages to try to improve identification performance. The prior method has obvious limitations that firstly, in the process of feature extraction of the traditional CNN model, the pooling operation is easy to lose the spatial hierarchy information of signals, and the single or simple combination convolution kernel is difficult to comprehensively capture the local details and global associated features of the RF time-frequency diagram, so that the feature characterization capability is insufficient, secondly, the single-layer convolution layer of the original capsule network is limited when processing the large-size radio frequency time-frequency diagram, the number of primary capsules is easy to be excessive, the network calculation efficiency is influenced, thirdly, the traditional Inception structure is matched with the capsule network model, no targeted optimization is carried out on Inception modules, and the parameter efficiency and multi-scale fusion precision of feature extraction still have room. In addition, the model is affected by signal path attenuation, environmental scattering and electromagnetic interference in a long-distance and low signal-to-noise ratio scene, the recognition accuracy is easy to be greatly reduced, and the model is difficult to adapt to a complex practical application scene. Therefore, there is an urgent need for an unmanned aerial vehicle radio frequency signal identification method to improve accuracy and stability of unmanned aerial vehicle radio frequency signal identification. Disclosure of Invention The application mainly aims to provide an unmanned aerial vehicle radio frequency signal identification method device, equipment and medium based on an improved capsule network model, and aims to solve the technical problem of how to improve the accuracy and stability of unmanned aerial vehicle radio frequency signal identification. In order to achieve the above purpose, the application provides an unmanned aerial vehicle radio frequency signal identification method based on an improved capsule network model, comprising the following steps: Collecting unmanned aerial vehicle radio frequency original signals with different distances; Performing short-time Fourier transform and logarithmic transformation on the unmanned aerial vehicle radio frequency original signal to obtain a radio frequency time-frequency diagram; And inputting the radio frequency time-frequency diagram into an improved capsule network model for feature extraction and classification to obtain an unmanned aerial vehicle identification result, wherein the improved capsule network model comprises a 7×7 convolution layer, an improved Inception feature extraction module, a primary capsule layer and a digital capsule layer, and the improved Inception feature extraction module comprises three improved Inception structures. In an embodiment, the step of inputting the radio frequency time-frequency diagram into an improved capsule network model to perform feature extraction and classification to obtain the unmanned aerial vehicle recognition result includes: Inputting the radio frequency time-frequency diagram into a 7 multiplied by 7 convolution layer of an improved capsule network model to perform preliminary feature extraction to obtain an initial feature diagram; processing the initial feature map through an improved Inception structure of the improved Inception f