Unmanned aerial vehicle radio frequency fingerprint extraction method based on multi-dimensional feature field coding

Abstract

The invention discloses an unmanned aerial vehicle radio frequency fingerprint extraction method based on multi-dimensional feature field coding, and belongs to the field of wireless signal identification. In order to solve the problem that a traditional time-frequency diagram mainly represents signal energy distribution and is difficult to explicitly express a signal state transition relation, the invention provides an F-FD-A (instantaneous frequency-frequency difference-instantaneous amplitude) multi-dimensional feature coding strategy which can effectively extract information such as a frequency hopping rule and envelope deformation of a signal. The method comprises the following steps: firstly, extracting instantaneous frequency, frequency difference and instantaneous amplitude of a radio frequency signal; further, the Markov transfer field is utilized to encode the first two sequences, and the Gramer angle difference field is utilized to encode the amplitude sequence; and finally, fusing the feature images into an RGB image, and inputting the RGB image into a ResNet-50 network for identification. Experiments show that the method has excellent recognition precision in a mixed data set covering DJI multi-series and multi-brand universal models, and fine-grained and high-robustness recognition of similar models of unmanned aerial vehicles in a complex environment is realized.

Description

Technical Field

[0001] This invention belongs to the field of radio frequency fingerprint recognition technology, specifically relating to a method for extracting radio frequency fingerprints from drones based on multidimensional feature field coding. Background Technology

[0002] With the rapid development of civilian drone technology, the regulation of low-altitude safety faces severe challenges. Drone communication signals typically employ technologies such as frequency hopping spread spectrum (FHSS) or orthogonal frequency division multiplexing (OFDM), exhibiting non-stationary and nonlinear characteristics in complex electromagnetic environments.

[0003] Existing RF fingerprinting methods largely rely on time-frequency analysis (such as STFT and CWT). While these methods can reveal energy distribution, they often struggle to explicitly characterize the transition probability relationships between signal states and neglect temporal correlations over long time series. Directly utilizing I / Q data for deep learning is feasible, but it suffers from high data dimensionality and extremely high hardware computational demands. Therefore, designing a method that can more effectively capture signal frequency hopping patterns, dynamic changes, and envelope features, and efficiently characterize RF fingerprints with lower dimensionality, is a pressing issue that needs to be addressed. Summary of the Invention

[0004] To address the problems mentioned in the background, this invention proposes a UAV radio frequency fingerprinting method based on multidimensional feature field coding and a deep residual network. The core of this invention lies in designing an "F-FD-A" (instantaneous frequency-frequency difference-instantaneous amplitude) multidimensional feature coding strategy, which maps a one-dimensional time series into a three-channel feature field image containing physical meaning, and combines this with an improved residual network to achieve high-precision end-to-end recognition. The specific technical solution is as follows:

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A multi-dimensional feature field collaborative coding strategy based on signal physical properties:

[0007] Step 1: Decouple the original I / Q data into three physical dimensions: instantaneous frequency (F), frequency difference (FD), and instantaneous amplitude (A). The instantaneous frequency sequence represents the carrier change trajectory of the signal; the frequency difference sequence, through first-order difference operations, highlights the abrupt edges of the signal in the time domain (such as frequency hopping moments) and eliminates constant term interference; the instantaneous amplitude sequence represents the power envelope fluctuations and modulation texture of the signal.

[0008] Step 2: Markov Transfer Field (MTF) Encoding of Instantaneous Frequency. Considering the discrete state switching characteristics of UAV frequency-hopping signals, the MTF algorithm is used to map the one-dimensional frequency sequence into a two-dimensional image. MTF explicitly encodes the time-dimensional frequency hopping logic by calculating the Markov transfer matrix between quantile intervals. The geometric structure in the MTF image (such as blocky or grid-like textures) directly reflects the frequency-hopping pattern of the UAV.

[0009] Step 3: Markov Transfer Field (MTF) Encoding of Frequency Differences. The frequency difference sequence is also mapped to an MTF image. Since the difference sequence is sensitive to the rate of change, this channel focuses on enhancing the statistical characteristics of the dynamic changes in the signal, effectively suppressing the influence of static frequency offset.

[0010] Step 4: Gram Difference Field (GADF) Encoding of Instantaneous Amplitudes. The GADF algorithm is used to encode the amplitude sequence. GADF maps the normalized amplitude to a polar coordinate system and calculates the correlation matrix using trigonometric difference angle formulas. While preserving the signal's time dependence, GADF amplifies subtle textures in the signal envelope (such as rising edge jitter and flat-top noise) through nonlinear transformation.

[0011] Step 5: Based on the principle of three primary colors (RGB) synthesis, the frequency feature matrix, frequency difference feature matrix, and amplitude feature matrix obtained in steps 2, 3, and 4 are synthesized into a "F-FD-A" three-channel feature field image containing multi-dimensional physical properties by taking the normalized frequency feature matrix as the R channel, the frequency difference feature matrix as the G channel, and the amplitude feature matrix as the B channel.

[0012] Step 6: Input the “F-FD-A” three-channel feature field image synthesized in Step 5 into the preset improved deep residual network for feature learning and classification, thereby achieving end-to-end high-precision identification of UAV radio frequency fingerprints.

[0013] Compared with the prior art, the beneficial effects of the present invention are:

[0014] This invention provides a new dimension for feature representation. By using Markov transition field (MTF) encoding, the analysis focuses on the transition probabilities of frequency sequences, which can be used to characterize macroscopic behaviors such as frequency hopping with discrete state switching. By using Gram angle difference field (GADF), the invention reveals the temporal correlation and microscopic texture in amplitude sequences that are not easily detected by traditional methods.

[0015] Existing technologies such as Short-Time Fourier Transform (STFT), when converting signals into time-frequency spectra, essentially represent the time-frequency distribution of signal energy. However, they have inherent limitations in characterizing signals with discrete state-hopping characteristics: firstly, due to the principle of time-frequency uncertainty, it is difficult to simultaneously obtain high frequency resolution at high time resolution, making it challenging to capture and finely characterize frequency-hopping transients; secondly, it is difficult to explicitly express the temporal transition patterns between signal states (such as frequency points). This often leads to the loss of crucial dynamic transition information and structural correlation features when traditional methods analyze signals with obvious state switching and temporal dependence, such as frequency hopping signals. In contrast, this invention, by employing Markov transition fields and Gram angle difference fields, can provide a more fundamental and structured characterization of signal hopping behavior and temporal correlation from both the macroscopic dynamics of state transitions and the microscopic structure of sequence changes, thereby overcoming the limitations of traditional energy representations in terms of information dimension and expressive power. Attached Figure Description

[0016] Figure 1 This is a flowchart of the drone radio frequency fingerprint recognition system of this application;

[0017] Figure 2 This is a schematic diagram of feature visualization and RGB fusion based on MTF and GADF proposed in this invention;

[0018] Figure 3 This is a comparison image of multiple samples with 3-channel features proposed in this invention;

[0019] Figure 4 This is a scatter plot of the class separability verification of fingerprint features based on t-SNE proposed in this invention;

[0020] Figure 5 It is a confusion matrix diagram used for classification on the ResNet50 network using the RGB fusion feature map of this invention. Detailed Implementation

[0021] To facilitate understanding of the technical content of this invention by those skilled in the art, the invention will be further described in detail below with reference to the accompanying drawings and specific examples. It should be understood that the specific examples described herein are merely illustrative and not intended to limit the scope of the invention.

[0022] Considering the characteristics of UAV communication signals, we note their core dynamic behaviors, such as the duration of frequency hopping signals and the switching intervals between different frequency blocks, which typically occur on a time scale of milliseconds. This means that to capture and analyze these macroscopic dynamic changes and state transition characteristics, it is not necessary to directly process I / Q data at an ultra-high sampling rate on the same order of magnitude as the original carrier frequency.

[0023] Implementation Examples

[0024] This embodiment provides a method for extracting radio frequency fingerprints of unmanned aerial vehicles (UAVs) based on multidimensional feature field coding. The specific implementation steps are shown in Figure 1, and the main process includes:

[0025] 1. Decoupling of signal preprocessing and physical features

[0026] First, for the acquired raw complex baseband signal sequence S(n)=I(n)+jQ(n) from UAV communication, the instantaneous physical features are extracted using the following formula:

[0027] Instantaneous amplitude (A): The energy envelope of the response signal.

[0028]

[0029] Instantaneous frequency (F): Reflects the dynamic shift of the carrier frequency.

[0030]

[0031] Frequency Differential (FD): The rate at which frequency changes are reflected, used to capture frequency hopping transients.

[0032]

[0033] 2. The principle of "F-FD-A" three-channel feature field encoding

[0034] R-channel and G-channel: Markov Transfer Field (MTF) Encoding

[0035] For the frequency sequence F and the difference sequence FD, MTF encoding is used to capture the state transition probabilities.

[0036] Binning: The range of eigenvalues ​​is divided into Q quantile intervals (Bins), each value point Assigned to the corresponding interval .

[0037] Transition Matrix Construction: Calculate the first-order Markov transition matrix W, where the elements are... Indicates the range is in The data points in the interval will jump to the next time step. The probability of the interval:

[0038]

[0039] Time dimension mapping: To preserve time correlation, the MTF matrix M is defined as follows:

[0040]

[0041] The resulting image (as shown in Figure 2) can explicitly show the hopping pattern of the frequency hopping point.

[0042] Channel B: Gram Angular Difference Field (GADF) Encoding

[0043] For amplitude sequence A, GADF encoding is used to preserve temporal dependence and amplify micro-texture.

[0044] Normalization: Scaling A to the [0,1] interval.

[0045] Polar coordinate mapping:

[0046]

[0047] Calculate the angular difference field: Construct the correlation matrix using the trigonometric function difference angle formula:

[0048]

[0049] 3. Fingerprint recognition model based on improved deep residual network

[0050] This embodiment uses an improved ResNet50 as the core recognition network. The experimental configuration and training steps are as follows:

[0051] Data augmentation: The Mixup data augmentation strategy is adopted to expand the decision boundary and enhance the robustness of the model against noise interference by linearly interpolating the input samples and their labels.

[0052] Classification head optimization: A Dropout layer (with a dropout rate of 0.3) is introduced after the global average pooling layer to suppress overfitting, followed by a fully connected layer to output the probability distribution of the corresponding drone category.

[0053] Loss function selection: In order to solve the potential imbalance problem in sample collection from different types of drones, this embodiment uses Focal Loss as the loss function. By introducing a weight factor and a focusing parameter (γ=2.0), the model pays more attention to difficult-to-classify samples during training.

[0054] Optimizer and Scheduling: The AdamW optimizer was used, with weight decay set to 0.05. The learning rate was dynamically adjusted using a warmup (5 epochs) combined with cosine annealing.

[0055] 4. Analysis of Experimental Results

[0056] Feature separability verification: As shown in Figure 4, the extracted "F-FD-A" feature field is visualized by dimensionality reduction using the t-SNE algorithm. It can be seen that the RF fingerprints of different UAV models exhibit a significant clustering effect in the feature space, with large inter-class distances, proving that the encoding strategy proposed in this invention can effectively extract RF fingerprints with high discriminative power.

[0057] Performance Evaluation: The model's performance on the test set is shown in the confusion matrix of Figure 5. Experimental data demonstrate that the method combining multidimensional feature field encoding and deep residual networks maintains high recognition accuracy even in complex low-altitude electromagnetic environments.

Claims

1. A method for radio frequency fingerprinting of unmanned aerial vehicles (UAVs) based on multidimensional feature field coding and deep residual networks, characterized in that, Includes the following steps: Step S1: Collect the radio frequency signal of the UAV communication link and obtain I / Q quadrature baseband data; Step S2: Preprocess and segment the I / Q data to extract a one-dimensional physical feature set containing instantaneous frequency sequence, frequency difference sequence and instantaneous amplitude sequence; Step S3: Employ the "F-FD-A" multidimensional feature encoding strategy to map the one-dimensional physical feature set into a two-dimensional feature image; wherein, using the Markov Transition Field (MTF) algorithm, based on the state transition probability matrix, the instantaneous frequency sequence is encoded into an R-channel image and the frequency difference sequence is encoded into a G-channel image; using the Gram Angular Difference Field (GADF) algorithm, based on trigonometric function transformation in polar coordinates, the instantaneous amplitude sequence is encoded into a B-channel image; Step S4: Fuse the images of the R, G, and B channels to construct an RGB three-channel radio frequency fingerprint image; Step S5: Construct a deep residual network ResNet-50 and introduce the Focal Loss loss function and the Mixup data augmentation strategy; Step S6: Train the improved deep residual network using the generated radio frequency fingerprint image dataset, input the signal to be tested into the trained network model, and output the identification results of the UAV model and operating conditions.

2. The method according to claim 1, characterized in that, The specific process for extracting the one-dimensional physical feature set in step S2 is as follows: Calculate complex signals instantaneous amplitude and instantaneous phase ; For instantaneous phase After unwinding, the phase difference is calculated to obtain the instantaneous frequency sequence. ; For instantaneous frequency sequences Perform first-order difference to obtain the frequency difference sequence. .

3. The method according to claim 1, characterized in that, The specific mathematical process of encoding using Markov Transition Fields (MTF) in step S3 includes: S3-1: Input one-dimensional time series The value range is divided into Q quantile intervals, and each time point Mapped to the corresponding range index ; S3-2: Calculate the first-order Markov transition matrix W, whose elements This indicates that in a time series, the values ​​range from an interval Then move to the interval The probability is given by the formula: ; S3-3: Construct the Markov transition field matrix M based on the time order of the sequence, and its matrix elements... Defined as the transition probability between the interval containing time step i and the interval containing time step j, i.e. ; The MTF image is generated by the M matrix, which explicitly depicts the discrete state transition law and frequency hopping dynamic characteristics of the UAV radio frequency signal in the frequency domain.

4. The method according to claim 1, characterized in that, The specific mathematical process of encoding using the Gram Angular Difference Field (GADF) in step S3 includes: S3-4: Convert the instantaneous amplitude sequence Perform min-max normalization to scale its numerical range to... , to obtain the sequence ; S3-5: Map the normalized sequence to the polar coordinate system using the inverse cosine function to obtain the angle sequence. The formula is ,in ; S3-6: Calculate the Gram angle difference field matrix G using the trigonometric function difference angle formula. Its elements are defined as follows: ; The GADF image is generated by matrix G, which preserves the temporal dependence of the signal and enhances the nonlinear texture features of the signal envelope through sinusoidal transformation.

5. The method according to claim 1, characterized in that, The Focal Loss function introduced in step S5 is used to solve the problems of class imbalance and difficult / easy sample mining in multi-class classification tasks. Its mathematical expression is: ,in, This represents the model's predicted probability of the true class. γ is the balance factor, and γ is the focusing parameter.

Images