CN-121997310-A - Mamba-based multi-domain feature fusion unmanned aerial vehicle method and device for identifying radio frequency fingerprint of machine

CN121997310ACN 121997310 ACN121997310 ACN 121997310ACN-121997310-A

Abstract

The invention discloses a Mamba-based multi-domain feature fusion unmanned aerial vehicle radio frequency fingerprint identification method and device, which relate to the technical field of wireless communication and target identification and comprise the steps of fusing in-phase/quadrature features, bispectral features and short-time Fourier spectral features of signals to realize the enhanced characterization of radio frequency fingerprints. Aiming at the computational complexity and real-time bottleneck brought by multi-domain feature fusion, a high-efficiency fusion framework based on Mamba and Vision-Mamba is constructed, the framework fully exerts the advantages of linear computational complexity of a Mamba model in long-sequence modeling, combines with a hardware-aware algorithm optimization strategy, and remarkably reduces computational overhead while maintaining high recognition precision. Aiming at the problem that redundancy and complementarity are difficult to balance in the feature fusion process, a self-adaptive feature interaction module based on a cross attention mechanism is introduced, and the self-adaptive feature interaction module can dynamically model the dependency relationship among multiple source features to realize self-adaptive distribution of fusion weights, thereby effectively inhibiting feature degradation phenomenon caused by environmental interference.

Inventors

GAO DAWEI
NIE GUANGYUAN
CHEN LIYANG
LIAO GUISHENG
CHEN YUFENG
Zeng Huiran
WANG YINSHENG

Assignees

西安电子科技大学杭州研究院
西安电子科技大学

Dates

Publication Date: 20260508
Application Date: 20251219

Claims (10)

1. Mamba-based multi-domain feature fusion unmanned aerial vehicle radio frequency fingerprint identification method is characterized by comprising the following steps: Acquiring an unmanned aerial vehicle radio frequency signal, representing the unmanned aerial vehicle radio frequency signal by adopting an in-phase component and a quadrature component, acquiring a bispectrum characteristic according to the unmanned aerial vehicle radio frequency signal, and acquiring a short-time Fourier transform characteristic according to the unmanned aerial vehicle radio frequency signal; The unmanned aerial vehicle radio frequency signal, the bispectrum feature and the short-time Fourier transform feature are processed through a trained network to obtain a classification result, the unmanned aerial vehicle radio frequency signal is processed through a first branch module of a Mamba-based multi-domain feature fusion network in the trained network to obtain a time domain feature, the bispectrum feature is processed through a second branch module of the Mamba-based multi-domain feature fusion network in the trained network to obtain a high-order statistical feature, the short-time Fourier transform feature is processed through a third branch module of the Mamba-based multi-domain feature fusion network in the trained network to obtain a frequency spectrum feature, the time domain feature, the high-order statistical feature and the frequency spectrum feature are fused through a cross attention module in the trained network to obtain a first fusion feature and a second fusion feature, the first fusion feature and the second fusion feature are fused to obtain a fusion feature, the fusion feature is classified through a classification head in the trained network to obtain the classification result, and the first fusion feature, the short-time Fourier transform feature and the short-time Fourier transform feature are fused.
2. The method for identifying the radio frequency fingerprint of the Mamba-based multi-domain feature fusion unmanned aerial vehicle according to claim 1, wherein the Mamba-based multi-domain feature fusion network in the trained network comprises a first branching module, a second branching module and a third branching module, The first branching module comprises a first normalization layer, a first learnable linear layer, a second learnable linear layer, a first convolution layer, a first nonlinear activation function, a second nonlinear activation function, a selective state space model and a global average pooling layer; The second branching module comprises a second convolution layer, a third nonlinear activation function, a first layer normalization layer, a second normalization layer, a third learnable linear layer, a fourth learnable linear layer, a first forward one-dimensional convolution layer, a fourth nonlinear activation function, a first forward state space model, a first backward one-dimensional convolution layer, a fifth nonlinear activation function, a first backward state space model, a sixth nonlinear activation function and a first linear layer; The third branching module comprises a third convolution layer, a seventh nonlinear activation function, a second layer normalization layer, a third normalization layer, a fifth learnable linear layer, a sixth learnable linear layer, a second forward one-dimensional convolution layer, an eighth nonlinear activation function, a second forward state space model, a second backward one-dimensional convolution layer, a ninth nonlinear activation function, a second backward state space model, a tenth nonlinear activation function and a second linear layer; the cross attention module in the trained network comprises a linear mapping feature embedding module, a first softmax function, a second softmax function, a first nonlinear transformation module and a second nonlinear transformation module; The classification head in the trained network comprises a third linear layer, a seventh nonlinear activation function, a regularization constraint layer and a classification layer.
3. The Mamba-based multi-domain feature fusion unmanned aerial vehicle radio frequency fingerprint identification method according to claim 2, wherein the processing of the unmanned aerial vehicle radio frequency signal by the first branch module of the Mamba-based multi-domain feature fusion network in the trained network to obtain the time domain feature comprises: Downsampling the unmanned aerial vehicle radio frequency signal to obtain an IQ sequence ; Normalizing the IQ sequence by adopting the first normalization layer to obtain a first characteristic; Processing the first feature by using the first and second leachable linear layers respectively, and projecting the processed first feature into two different feature spaces to obtain a second feature And third feature Expressed as: ; ; Wherein, the The feature mapping module is represented as a function of the feature, The real number domain is represented by the number, Representing the length of the input feature map, Representing the width of the input feature map; the first convolution layer is adopted to carry out convolution processing on the second feature, the first nonlinear activation function is adopted to process the feature after the convolution processing to obtain a fourth feature, and the selective state space model is adopted to process the fourth feature to obtain a fifth feature Expressed as: ; Wherein, the A state space model is represented and is used to represent a state space model, The activation function is represented as a function of the activation, Representing a one-dimensional convolution; and processing the third feature by adopting the second nonlinear activation function to obtain a sixth feature, wherein the sixth feature is expressed as: ; Fusing the fifth feature and the sixth feature in a way of multiplying the fifth feature and the sixth feature element by element to obtain a seventh feature, wherein the seventh feature is expressed as follows: ; Wherein, the Representing element-by-element multiplication; adding the seventh feature and the IQ sequence to obtain an eighth feature, and processing the eighth feature by adopting the global average pooling layer to obtain a time domain feature Expressed as: ; Wherein, the An index representing the input feature map.
4. The Mamba-based multi-domain feature fusion unmanned aerial vehicle radio frequency fingerprint identification method according to claim 2, wherein the processing the bispectral features through a second branch module of a Mamba-based multi-domain feature fusion network in the trained network to obtain high-order statistical features comprises: Converting the bispectral features into amplitude spectrum features, dividing the amplitude spectrum features into a plurality of patches, carrying out convolution processing on the plurality of patches by adopting the second convolution layer, and processing the features after the convolution processing by adopting the third nonlinear activation function so as to realize image block embedding, thereby obtaining embedded features corresponding to the patches, wherein the embedded features are expressed as follows: ; Wherein, the Representing a feature block in which the amplitude spectrum features are divided, The representation SiLU activates the function, Representing the convolution kernel that can be learned, The characteristics of the amplitude spectrum are represented, The number of output channels is indicated and, , , The downsampling factor is represented as such, Representing the width of the amplitude spectrum feature, A height representing an amplitude spectrum feature; Flattening the embedded features corresponding to each Patch in the spatial dimension, and injecting a learnable two-dimensional position code , And carrying out layer normalization processing by adopting the first layer normalization layer to obtain a ninth feature Expressed as: ; Wherein, the A layer normalization operation is represented and is performed, Representing a flattening operation; Normalizing the ninth feature by adopting the second normalization layer to obtain a tenth feature; Processing the tenth feature by using the third and fourth learnable linear layers respectively, and projecting the processed tenth feature into two different feature spaces to obtain an eleventh feature And twelfth feature Expressed as: ; ; Wherein, the And Representing a different linear layer mapping of the layer, Representing a normalization operation; The eleventh feature is convolved by adopting the first forward one-dimensional convolution layer, the feature after the convolution processing is processed by adopting the fourth nonlinear activation function to obtain a thirteenth feature, and the thirteenth feature is processed by adopting the first forward state space model to obtain a fourteenth feature Expressed as: ; Wherein, the Representing the processing procedure of the forward state space model; The eleventh feature is convolved by adopting the first backward one-dimensional convolution layer, the feature after the convolution processing is processed by adopting the fifth nonlinear activation function to obtain a fifteenth feature, and the fifteenth feature is processed by adopting the first backward state space model to obtain a sixteenth feature Expressed as: ; Wherein, the A process of representing a backward state space model; Processing the twelfth feature by adopting the sixth nonlinear activation function to obtain a seventeenth feature, and fusing the seventeenth feature and the fourteenth feature in a way of multiplying the seventeenth feature and the fourteenth feature element by element to obtain an eighteenth feature, wherein the eighteenth feature is expressed as: ; Fusing the seventeenth feature and the sixteenth feature by multiplying the seventeenth feature element by element to obtain a nineteenth feature, expressed as: ; Adding the eighteenth feature to the nineteenth feature to obtain a twentieth feature Expressed as: ; processing the twentieth feature by adopting the first linear layer, and adding the processed feature and the ninth feature to obtain the high-order statistical feature 。
5. The Mamba-based multi-domain feature fusion unmanned aerial vehicle radio frequency fingerprint identification method according to claim 2, wherein the processing the short-time fourier transform feature by the third branch module of the Mamba-based multi-domain feature fusion network in the trained network to obtain a spectrum feature comprises: dividing the short-time Fourier transform feature into a plurality of patches, carrying out convolution processing on the plurality of patches by adopting the third convolution layer, and processing the feature after the convolution processing by adopting the seventh nonlinear activation function so as to realize image block embedding, thereby obtaining embedded features corresponding to the patches, wherein the embedded features are expressed as follows: ; Wherein, the A feature map representing the short-time fourier transform features divided into small blocks, The representation SiLU activates the function, Representing the convolution kernel that can be learned, Representing the features of the short-time fourier spectrum, The number of output channels is indicated and, , , The downsampling factor is represented as such, Representing the width of the short-time fourier spectrum features, A height representing a short-time fourier spectrum characteristic; Flattening the embedded features corresponding to each Patch in the spatial dimension, and injecting a learnable two-dimensional position code , And carrying out layer normalization processing by adopting the second layer normalization layer to obtain a twenty-first characteristic Expressed as: ; Wherein, the A layer normalization operation is represented and is performed, Representing a flattening operation; normalizing the twenty-first features by adopting the third normalization layer to obtain twenty-second features; Processing the twenty-second feature by using the fifth and sixth learnable linear layers respectively, and projecting the processed twenty-second feature into two different feature spaces to obtain twenty-third feature And twenty-fourth feature Expressed as: ; ; Wherein, the And Representing a different linear layer mapping of the layer, Representing a normalization operation; The twenty-third feature is convolved by adopting the second forward one-dimensional convolution layer, the feature after the convolution processing is processed by adopting the eighth nonlinear activation function to obtain a twenty-fifth feature, and the twenty-fifth feature is processed by adopting the second forward state space model to obtain a twenty-sixth feature Expressed as: ; Wherein, the Representing the processing procedure of the forward state space model; the twenty-third feature is convolved by adopting the second backward one-dimensional convolution layer, the feature after the convolution processing is processed by adopting the ninth nonlinear activation function to obtain a twenty-seventh feature, and the twenty-seventh feature is processed by adopting the second backward state space model to obtain a twenty-eighth feature Expressed as: ; Wherein, the A process of representing a backward state space model; The twenty-fourth feature is processed by adopting the tenth nonlinear activation function to obtain a twenty-ninth feature, and the twenty-ninth feature and the twenty-sixth feature are fused in a way of multiplying the twenty-sixth feature element by element to obtain a thirty-fourth feature, which is expressed as: ; fusing the twenty-ninth feature and the twenty-eighth feature by multiplying the twenty-eighth feature element by element to obtain a thirty-first feature, which is expressed as: ; Adding the thirty-first feature to the thirty-second feature to obtain a thirty-second feature Expressed as: ; Processing the thirty-second feature by adopting the second linear layer, and adding the processed feature and the twenty-first feature to obtain the frequency spectrum feature 。
6. The Mamba-based multi-domain feature fusion unmanned aerial vehicle radio frequency fingerprint identification method according to claim 2, wherein the time domain feature, the high-order statistical feature and the spectrum feature are fused through a cross attention module in a trained network to obtain a first fusion feature and a second fusion feature, and the first fusion feature and the second fusion feature are fused to obtain a fusion feature, comprising: the time domain features are subjected to the feature embedding module Said high order statistical features And the spectral features The embedding process is performed as follows: ; ; ; Wherein, the 、 And Representing projection of temporal features onto queries, respectively Key and key Sum value The vector of the back-up vector is calculated, 、 And Representing projected temporal features The corresponding weight to be able to learn is, 、 And Representing spectral features projected onto queries separately Key and key Sum value The vector of the back-up vector is calculated, 、 And Representing projected spectral features The corresponding weight to be able to learn is, 、 And Representing projection of higher-order statistical features onto queries, respectively Key and key Sum value The vector of the back-up vector is calculated, 、 And Representing projected higher order statistical features Corresponding learnable weights; Calculating a time domain attention moment array by adopting a cross attention mechanism and combining the first softmax function and the second softmax function And Expressed as: ; ; Wherein, the Representation of Is used in the manufacture of a printed circuit board, Representation of Is used in the manufacture of a printed circuit board, Representing a matrix transpose; attention moment array for time domain by adopting the first nonlinear transformation module Processing, processing the result and the high-order statistical feature Adding to obtain a first fusion feature Expressed as: ; The second nonlinear transformation module is adopted to pay attention to the moment array Processing, processing the result and the high-order statistical feature Adding to obtain a second fusion feature Expressed as: ; Wherein, the Representing a nonlinear transformation function; Characterizing the first fusion feature And the second fusion feature Fusing to obtain fusion characteristics, wherein the fusion characteristics are expressed as follows: ; Wherein, the Representing the fusion factor.
7. The Mamba-based multi-domain feature fusion unmanned aerial vehicle radio frequency fingerprint identification method according to claim 1, wherein the training process of the trained network comprises the following steps: acquiring preset class data as a sample, and acquiring a real label of the sample to construct a training data set, wherein the sample comprises unmanned aerial vehicle radio frequency signals, bispectral features and short-time Fourier transform features; inputting a portion of the samples in the training dataset to a first Training the network to be trained for the second time to obtain the first time The prediction result output by the classification head in the secondary training process; According to the first Prediction result output by classification head in secondary training process and training first The real labels of the samples of the network to be trained are used for calculating the classification loss and serving as the first step Classification loss in the secondary training process; According to the first The classification loss of the secondary training process is back-propagated to update the first Obtaining network parameters of the network to be trained for the second time to obtain the first time And iterating until the training times or the convergence degree meet the preset conditions to obtain the trained network.
8. The Mamba-based multi-domain feature fusion unmanned aerial vehicle radio frequency fingerprint identification method according to claim 7, wherein the expression for calculating the classification loss is: ; Wherein, the Representing the Focal Loss function, Representing true categories Is used to determine the prediction probability of (1), Representing the weighting coefficients for handling the class imbalance, Representing the focus parameter.
9. The Mamba-based multi-domain feature fusion unmanned aerial vehicle radio frequency fingerprint identification method according to claim 1, wherein the unmanned aerial vehicle radio frequency signal Expressed as: ; Wherein, the Representing the in-phase component, Representing the orthogonal component(s), Representing imaginary units; According to the unmanned aerial vehicle radio frequency signal, calculating the third-order cumulative quantity of the unmanned aerial vehicle radio frequency signal Expressed as: ; Wherein, the And Representing the different time delays of the time delay, Representing a desired operator; performing two-dimensional Fourier transform on the third-order cumulant to obtain bispectral features Expressed as: ; on a time axis, segmenting the unmanned aerial vehicle radio frequency signal, and carrying out Fourier transform on each segment to obtain a frequency spectrum of the signal changing along with time, namely a short-time Fourier transform characteristic, wherein the frequency spectrum is expressed as: ; Wherein, the The length of time is indicated and, Which represents the angular frequency of the light emitted by the light source, Indicating a time length of Is a function of the sliding window of (a), Time is indicated.
10. Mamba-based multi-domain feature fusion unmanned aerial vehicle radio frequency fingerprint identification device is characterized by comprising: the device comprises a feature acquisition module, a double-spectrum feature acquisition module, a short-time Fourier transform feature acquisition module and a short-time Fourier transform module, wherein the feature acquisition module is used for acquiring an unmanned aerial vehicle radio frequency signal and representing the unmanned aerial vehicle radio frequency signal by adopting an in-phase component and a quadrature component; The feature processing fusion module is used for processing the unmanned aerial vehicle radio frequency signals, the bispectrum features and the short-time Fourier transform features by adopting a trained network to obtain classification results, processing the unmanned aerial vehicle radio frequency signals by a first branch module of a Mamba-based multi-domain feature fusion network in the trained network to obtain time domain features, processing the bispectrum features by a second branch module of the Mamba-based multi-domain feature fusion network in the trained network to obtain high-order statistical features, processing the short-time Fourier transform features by a third branch module of the Mamba-based multi-domain feature fusion network in the trained network to obtain frequency spectrum features, fusing the time domain features, the high-order statistical features and the frequency spectrum features by a cross attention module in the trained network to obtain first fused features and second fused features, fusing the first fused features and the second fused features to obtain fusion features, classifying the fusion features by a classification head in the trained network to obtain classification results, wherein the first short-time feature and the short-time Fourier transform features are fused by the first fused features and the short-time Fourier transform features.

Description

Mamba-based multi-domain feature fusion unmanned aerial vehicle method and device for identifying radio frequency fingerprint of machine Technical Field The invention belongs to the technical field of wireless communication and target identification, and particularly relates to a Mamba-based multi-domain feature fusion unmanned aerial vehicle radio frequency fingerprint identification method and device. Background With the rapid development of low-altitude economy and unmanned systems, unmanned aerial vehicles are widely applied in the fields of logistics transportation, environmental monitoring, security inspection and the like. However, a large number of "black flying" unmanned aerial vehicles frequently appear in the low-altitude airspace, which causes electromagnetic interference, information leakage and public safety hidden trouble, and brings serious challenges to the control of the low-altitude airspace. Traditional unmanned aerial vehicle body recognition relies on communication protocol analysis recognition, and reliable recognition of non-cooperative targets is difficult in a complex electromagnetic environment. The current non-cooperative unmanned aerial vehicle target recognition research mainly focuses on vision and radar sensors, although certain progress is made in the aspects of target detection and model recognition, obvious limitations still exist for individual recognition of the non-cooperative unmanned aerial vehicle target, the existing method is mostly based on vision analysis or expert system rules, relies on macroscopic statistical characteristics of signals, and small physical differences caused by hardware non-ideal characteristics are ignored, so that individual level distinction is insufficient. Accordingly, there is a need to provide a method for unmanned aerial vehicle identification that ameliorates the above-described deficiencies in the prior art. Disclosure of Invention In order to solve the problems in the prior art, the invention provides a Mamba-based multi-domain feature fusion unmanned aerial vehicle radio frequency fingerprint identification method and device. The technical problems to be solved by the invention are realized by the following technical scheme: in a first aspect, the invention provides a multi-domain feature fusion unmanned aerial vehicle radio frequency fingerprint identification method based on Mamba, which comprises the following steps: acquiring an unmanned aerial vehicle radio frequency signal, representing the unmanned aerial vehicle radio frequency signal by adopting an in-phase component and a quadrature component, acquiring a bispectrum characteristic according to the unmanned aerial vehicle radio frequency signal, and acquiring a short-time Fourier transform characteristic according to the unmanned aerial vehicle radio frequency signal; The method comprises the steps of processing unmanned aerial vehicle radio frequency signals, bispectrum features and short-time Fourier transform features by a trained network to obtain classification results, processing unmanned aerial vehicle radio frequency signals by a first branch module of a Mamba-based multi-domain feature fusion network in the trained network to obtain time domain features, processing bispectrum features by a second branch module of the Mamba-based multi-domain feature fusion network in the trained network to obtain high-order statistical features, processing short-time Fourier transform features by a third branch module of the Mamba-based multi-domain feature fusion network in the trained network to obtain frequency spectrum features, fusing the time domain features, the high-order statistical features and the frequency spectrum features by a cross attention module in the trained network to obtain first fused features and second fused features, fusing the first fused features and the second fused features to obtain fused features, and classifying the fused features by a classification head in the trained network to obtain classification results, wherein the first fused features fuse unmanned aerial vehicle radio frequency signals and the short-time Fourier transform features and the second fused features. In a second aspect, the present invention further provides a multi-domain feature fusion unmanned aerial vehicle radio frequency fingerprint identification device based on Mamba, including: The system comprises a feature acquisition module, a double-spectrum feature acquisition module, a short-time Fourier transform feature acquisition module and a short-time Fourier transform module, wherein the feature acquisition module is used for acquiring an unmanned aerial vehicle radio frequency signal and adopting an in-phase component and a quadrature component to represent the unmanned aerial vehicle radio frequency signal; The feature processing fusion module is used for processing the unmanned aerial vehicle radio frequency signals, the bispectrum features and the short-time Fourier tra