A method for extracting and identifying visual features of radio frequency signals of a drone
By employing VMD and Pearson correlation coefficient filtering for noise reduction, combined with Gram angular field matrix transformation, radio frequency signals are converted into two-dimensional images. Convolutional neural networks are then used for UAV type recognition, solving the problem of feature instability under complex interference and achieving high-accuracy UAV recognition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SICHUAN NORMAL UNIV
- Filing Date
- 2026-04-10
- Publication Date
- 2026-07-21
AI Technical Summary
Existing drone radio frequency identification technology suffers from unstable features and insufficient noise resistance under complex interference and low signal-to-noise ratio conditions, resulting in insufficient identification accuracy and robustness.
Variational mode decomposition (VMD) and Pearson correlation coefficient are used for noise reduction. Combined with Gram angular field matrix transformation, one-dimensional radio frequency signals are mapped into two-dimensional images, and lightweight convolutional neural networks are used for feature learning and classification.
It significantly improves signal-to-noise ratio and feature stability under low signal-to-noise ratio conditions, and improves the accuracy of drone type recognition, especially with a recognition rate of over 90% when the signal-to-noise ratio is 0dB.
Smart Images
Figure CN122432730A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radio signal processing technology, and in particular relates to a method for extracting and recognizing the visual features of radio frequency signals from unmanned aerial vehicles (UAVs). Background Technology
[0002] In recent years, drones have rapidly penetrated into scenarios such as emergency rescue, aerial surveying and mapping, agricultural plant protection, and public services. Their ease of operation and high mobility, while improving operational efficiency, also make them more susceptible to unauthorized activities such as crossing boundaries, entering no-fly zones, and disrupting public order. In actual supervision, relying solely on manual patrols or post-incident evidence collection is insufficient to meet the requirements of "rapid detection—timely judgment—precise handling" of drones, especially in complex electromagnetic environments and urban environments with multiple obstructions. How to reliably acquire key information for identification has become one of the core issues in drone management. Therefore, for airspace safety management, there is an urgent need for a technical solution capable of reliably extracting drone communication radio frequency signals under interference backgrounds and achieving type identification and target discrimination.
[0003] Existing methods for UAV target detection and identification typically include radar detection, visible / infrared detection, acoustic detection, and radio frequency (RF) detection. Visible / infrared detection relies on imaging quality and visibility conditions, and is easily affected by changes in lighting, weather attenuation, obstruction, and field-of-view limitations. Acoustic detection has some usability at close range, but is prone to false detections and missed detections under traffic noise, wind noise, and multiple sound source superposition conditions. In contrast, RF detection can directly intercept remote control or data transmission links between the UAV and the ground station from the airspace, is not dependent on visibility conditions, and is to some extent insensitive to obstruction, thus possessing strong applicability in engineering deployments.
[0004] In RFID research, existing methods typically rely on information such as time-domain waveforms, spectral amplitude, instantaneous features, or time-frequency graphs for discrimination, with some works employing deep learning to directly classify the spectrograms. However, in real-world applications, RF links are often affected by factors such as broadband noise, co-channel communication devices, adjacent-channel interference, and multipath fading, leading to significant non-stationary characteristics in the original sampled sequences. On the one hand, noise and interference can mask the discriminative fine-grained structures in UAV communication, resulting in insufficient stability of features based on a single representation. On the other hand, traditional manual feature extraction often relies on empirical selection, and the feature distribution is prone to drift under different models, different transmit powers, and different signal-to-noise ratios, causing fluctuations in recognition performance.
[0005] Furthermore, some RFID methods based on image representation typically convert sequences directly into two-dimensional images. However, if noise and non-target components in the signal are not effectively suppressed, the texture structure in the image will be severely contaminated by random perturbations, making it difficult for deep networks to learn stable and generalizable discrimination patterns. At the same time, if the image representation process fails to preserve the overall correlation between time-series points, it will weaken the ability to express the inherent temporal coupling characteristics of UAV signals, limiting the recognition upper limit of subsequent classifiers.
[0006] Therefore, in order to solve the problems of unstable radio frequency features and insufficient representation ability under interference and low signal-to-noise ratio conditions in the existing technology, this paper proposes a method for visual feature extraction and recognition of UAV radio frequency signals to solve the problems existing in the existing technology. Summary of the Invention
[0007] The purpose of this invention is to provide a method for extracting and recognizing visual features of UAV radio frequency signals, which solves the problems of insufficient noise resistance and inadequate feature representation of the original radio frequency time sequence signals under complex interference and low signal-to-noise ratio conditions in existing radio detection and recognition methods, resulting in insufficient recognition accuracy and robustness.
[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for extracting and recognizing visual features of UAV radio frequency signals includes the following steps: Step 1: Use a general-purpose software radio receiver to intercept the radio frequency communication link in the target airspace and collect the radio frequency signal between the UAV and the ground remote controller; Step 2: Divide the radio frequency signal into segments of 1024 sampling points each to obtain the set of original signal sample sequences to be processed; Step 3: Perform variational mode decomposition on the original signal sample sequence to obtain a series of mode components; at the same time, calculate the Pearson correlation coefficient between each mode component and the original signal sample sequence. Step 4: Use the mean of all Pearson correlation coefficients obtained in Step 3 as a threshold, compare the Pearson correlation coefficient of each modal component with the threshold to obtain the index set of effective modal components; superimpose and reconstruct the modal components in the index set to obtain the denoised radio frequency signal. Step 5: Normalize the denoised RF signal and perform polar coordinate transformation to obtain the corresponding angle sequence; construct the Gram angle field matrix based on the angle sequence, and map the one-dimensional RF signal into a two-dimensional image through the Gram angle field matrix; Step 6: Input the two-dimensional image obtained in Step 5 into the convolutional neural network for feature learning and classification decision, and output the drone type recognition result.
[0009] Furthermore, in step three, the specific calculation method for the Pearson correlation coefficient is as follows: For the original signal sample sequence Variational mode decomposition into K modal components can be expressed as: in, For the k-th time-domain mode component, For the k-th frequency domain spectral mode, The center frequency of the k-th mode; Subsequently, the Pearson correlation coefficient between each modal component and the original signal sample sequence was calculated. The formula for calculating its Pearson correlation coefficient is: in, express and covariance, express The variance of , where k represents the k-th modal component.
[0010] Furthermore, the method for obtaining the index set in step four is as follows: Based on all the correlation coefficients calculated in step three The mean value is used as the preset screening threshold. , where the threshold r is a dynamic threshold that is adaptively updated with the sample; The Pearson correlation coefficients of each modal component were then used. Compare with the screening threshold r; when When, determine that the k-th modal component is a valid modal component and retain it; when At that time, the k-th modal component is determined to be a noise-dominant component and is removed; thus, an index set of effective modal components is constructed. .
[0011] Furthermore, the radio frequency signal in step four is: .
[0012] Furthermore, in step five, we assume that the one-dimensional time series of the denoised RF signal is as follows: in, For the sequence length; for the sequence Perform normalization and scale it to an interval. The normalized sequence is obtained. ; For normalized sequences Perform polar coordinate transformation to obtain angle sequence ; Normalized sequence With the sequence of inverse cosine values Introducing polar coordinates to construct a polar radius term ; Based on angle sequence Construct the Gram angle field matrix, which consists of the Gram angle sum field and the Gram angle difference field: By using a two-dimensional matrix of Gram angle sum field GASF and Gram angle difference field GADF, one-dimensional radio frequency signals can be encoded into a two-dimensional image that can effectively characterize the correlation between different time points.
[0013] Furthermore, the formula for calculating the normalized sequence is: in, and Sequences The maximum and minimum values.
[0014] Furthermore, the polar coordinate transformation is implemented using the inverse cosine function, and its calculation formula is as follows: in Indicates sample The value of the inverse cosine.
[0015] Furthermore, the convolutional neural network in step six includes: four convolutional layers. The first convolutional layer has 32 kernels and a kernel size of 3×3. The second, third, and fourth convolutional layers have 64, 128, and 256 kernels, respectively, and each has a convolutional size of 3×3. Each convolutional layer is followed by a pooling layer with a pooling stride of 2. The convolutional output is flattened and connected to a fully connected layer with 512 hidden nodes. A Dropout layer with a Dropout ratio of 0.5 is then set after the fully connected layer. Finally, a Softmax classification layer is set to output the probabilities of different drone categories, and the category with the highest probability is used as the drone type recognition result.
[0016] Furthermore, in the training process of the convolutional neural network, the cross-entropy loss function is used as the optimization objective. The network weight parameters are updated through backpropagation until the loss function converges or reaches the preset number of training rounds, thereby obtaining a trained convolutional neural network classification model.
[0017] The beneficial effects of this invention are as follows: 1. This application employs Variational Mode Decomposition (VMD) to adaptively decompose the original radio frequency (RF) signal, obtaining multiple mode components. By calculating the Pearson correlation coefficient between each component and the original signal, and using the mean of all correlation coefficients as a dynamic threshold, highly correlated effective mode components are retained while noise-dominant components are removed. Finally, the retained components are superimposed to reconstruct the denoised RF signal. This scheme can adaptively separate signal and noise without relying on prior knowledge, effectively suppressing adverse effects such as broadband noise, co-channel interference, and multipath fading, thereby significantly improving the signal-to-noise ratio and characteristic stability of the RF signal under low signal-to-noise ratio conditions.
[0018] 2. This application does not rely on manually designed time-domain, frequency-domain, or instantaneous features. Instead, it constructs a Gram Angular Field Matrix (GASF / GADF) by normalizing and polar-coordinate transforming the denoised signal, encoding the one-dimensional radio frequency signal into a two-dimensional image. This graphical representation can automatically preserve the overall correlation and nonlinear coupling structure between time-series points and is unaffected by changes in specific signal parameters (such as model and transmit power), thereby avoiding the problem of feature distribution drift with the environment and enhancing the model's generalization ability under different signal-to-noise ratios and UAV types.
[0019] 3. This application employs rigorous adaptive denoising through VMD and Pearson correlation coefficient screening to ensure that the one-dimensional signal source of the input image has low noise interference. This results in the generated two-dimensional image showing clear and differentiated pattern structures for signals from different UAV types, rather than messy textures obscured by random perturbations. Therefore, the subsequent convolutional neural network can learn stable and distinguishable discriminative features, avoiding performance degradation caused by noise contamination.
[0020] 4. This application combines adaptive decomposition denoising (VMD and Pearson thresholding) with Gram angle field image transformation and employs a lightweight convolutional neural network for classification. In a specific implementation, simulation results show that even under extremely low signal-to-noise ratio (SNR) conditions of 0dB, the recognition rate of this method still exceeds 90%, significantly outperforming existing methods. This demonstrates that the proposed solution can effectively cope with complex electromagnetic environments and maintain highly reliable UAV type recognition capabilities even under strong interference. Attached Figure Description
[0021] Figure 1 This is a schematic diagram of the drone detection method of the present invention; Figure 2 The lightweight convolutional neural network model designed for this invention; Figure 3 This invention provides a visual representation of the intercepted UAV signals. Figure 4The results show the performance comparison of the UAV radio frequency signal visual feature extraction and recognition method provided by this invention. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0023] The application principle of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0024] See Figure 1 , Figure 2 , Figure 3 This embodiment provides a method for extracting and recognizing visual features of UAV radio frequency signals, including the following steps: Step 1: First, configure the radio platform according to the actual signal acquisition requirements, and use a general software radio receiver to intercept the radio frequency communication link in the target airspace to collect the radio frequency signal between the UAV and the ground remote controller.
[0025] Step 2: Divide the radio frequency signal into segments of 1024 sampling points each to obtain the set of original signal sample sequences to be processed.
[0026] Step 3: Perform variational mode decomposition on each original signal sample sequence to obtain a series of mode components; simultaneously calculate the Pearson correlation coefficient between each mode component and the original signal sample sequence. The specific calculation method for the Pearson correlation coefficient is as follows:
[0027] For the original signal sample sequence Variational mode decomposition into K modal components can be expressed as: in, For the k-th time-domain mode component, For the k-th frequency domain spectral mode, Let be the center frequency of the k-th mode.
[0028] Subsequently, the Pearson correlation coefficient between each modal component and the original signal sample sequence was calculated. The formula for calculating its Pearson correlation coefficient is: in, express and covariance, express The variance of , where k represents the k-th modal component.
[0029] Step 4: Analyze all Pearson correlation coefficients obtained in Step 3. The mean value is used as the threshold r, and the Pearson correlation coefficients of each modal component are compared with the threshold, i.e.: Wherein, the threshold r is a dynamic threshold that is adaptively updated with the samples. Then, the Pearson correlation coefficients of each modal component are... Compare with the threshold r. When When, determine that the k-th modal component is a valid modal component and retain it; when At that time, the k-th modal component is determined to be the dominant noise component and is removed; thus forming the effective modal component index set. The modal components in the index set are superimposed and reconstructed to obtain the denoised radio frequency signal. Its expression is:
[0030] Step 5: Normalize the denoised RF signal obtained in Step 4. Assume the one-dimensional time series of the denoised RF signal is as follows:
[0031] in, This is the sequence length. Then, the sequence... Perform normalization and scale it to an interval. The normalized sequence is obtained. The calculation formula is as follows:
[0032] in, and Sequences The maximum and minimum values. Then, the normalized sequence... Perform polar coordinate transformation to obtain angle sequence Polar coordinate transformation is achieved using the inverse cosine function, and its calculation formula is as follows:
[0033] in Indicates sample The inverse cosine value. Then the normalized sequence is... With the sequence of inverse cosine values By introducing polar coordinates, the ratio of the index value of the nth data point to the sequence length N is used as the radius of the nth data point in the polar coordinate system, thus constructing the polar coordinate radius term. (When converting to polar coordinates, only the angle and radius values can determine the position of a point.) The calculation formula is as follows:
[0034] Where n is the index value. Finally, based on the angle sequence... The Gram angle field matrix (GAF) is constructed, which is divided into the Gram angle sum field (GASF) and the Gram angle difference field (GADF). The calculation formulas are distributed as follows:
[0035] The one-dimensional radio frequency signal is then encoded into a two-dimensional image that can effectively characterize the correlation between different time points using the aforementioned GASF and GADF two-dimensional matrices.
[0036] Step Six: Divide the two-dimensional image samples obtained in Step Five into training data and test data, and input the training data into a convolutional neural network for parameter learning. The structure of the convolutional neural network is set to four convolutional layers. The first convolutional layer has 32 kernels and a kernel size of 3×3. The second, third, and fourth convolutional layers have 64, 128, and 256 kernels respectively, and the convolutional size is 3×3 for each layer. A pooling layer is set after each convolutional layer with a pooling stride of 2. The convolutional output is flattened and connected to a fully connected layer with 512 hidden nodes. A Dropout layer with a Dropout ratio of 0.5 is set after the fully connected layer. Finally, a Softmax classification layer is set to output the probability of different drone categories, and the category with the highest probability is used as the drone type recognition result. During training, the cross-entropy loss function is used as the optimization objective. The network weight parameters are updated through backpropagation until the loss function converges or the preset number of training rounds is reached, thereby obtaining the trained convolutional neural network classification model.
[0037] The visual features of UAV radio frequency signals are extracted and recognized based on the trained convolutional neural network classification model.
[0038] Compared to existing methods, this method eliminates the influence of environmental noise by adaptively decomposing the original signal and selecting the main modal components for signal reconstruction. It then maps the denoised one-dimensional sequence signal into a two-dimensional image using Gram Angular Field (GAF) transform, achieving signal visualization while amplifying the differences between signals from different UAV types, thus facilitating the identification of different UAVs. Figure 3 The signal visualization results show that the radio frequency signals of different drones exhibit different pattern shapes after Gram angle field transformation, which is more conducive to the neural network model learning the signal characteristics of different drones from two-dimensional images; from Figure 4The simulation results show that the proposed VMD-GASF method outperforms existing methods in overall UAV recognition performance under different signal-to-noise ratios, especially at low signal-to-noise ratios, such as when the signal-to-noise ratio SNR=0dB, the recognition rate of the proposed method exceeds 90%.
[0039] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for extracting and recognizing visual features of UAV radio frequency signals, characterized in that, Includes the following steps: Step 1: Use a general-purpose software radio receiver to intercept the radio frequency communication link in the target airspace and collect the radio frequency signal between the UAV and the ground remote controller; Step 2: Divide the acquired radio frequency signal into segments of 1024 sampling points each to obtain the set of original signal sample sequences to be processed; Step 3: Perform variational mode decomposition on the original signal sample sequence to obtain a series of mode components; at the same time, calculate the Pearson correlation coefficient between each mode component and the original signal sample sequence. Step 4: Use the mean of all Pearson correlation coefficients obtained in Step 3 as a threshold, compare the Pearson correlation coefficient of each modal component with the threshold to obtain the index set of effective modal components; superimpose and reconstruct the modal components in the index set to obtain the denoised radio frequency signal. Step 5: Normalize the denoised RF signal and perform polar coordinate transformation to obtain the corresponding angle sequence; construct the Gram angle field matrix based on the angle sequence, and map the one-dimensional RF signal into a two-dimensional image through the Gram angle field matrix; Step 6: Input the two-dimensional image obtained in Step 5 into the convolutional neural network for feature learning and classification decision, and output the drone type recognition result.
2. The method for extracting and recognizing visual features of UAV radio frequency signals according to claim 1, characterized in that: In step three, the specific calculation method for the Pearson correlation coefficient is as follows: For the original signal sample sequence Variational mode decomposition into K modal components can be expressed as: in, For the k-th time-domain mode component, For the k-th frequency domain spectral mode, The center frequency of the k-th mode; Subsequently, the Pearson correlation coefficient between each modal component and the original signal sample sequence was calculated. The formula for calculating its Pearson correlation coefficient is: in, express and covariance, express The variance of , where k represents the k-th modal component.
3. The method for extracting and recognizing visual features of UAV radio frequency signals according to claim 1, characterized in that: The method for obtaining the index set in step four is as follows: Based on all the correlation coefficients calculated in step three The mean value is used as the preset screening threshold. , where the threshold r is a dynamic threshold that is adaptively updated with the sample; The Pearson correlation coefficients of each modal component were then used. Compare with the screening threshold r; when When, determine that the k-th modal component is a valid modal component and retain it; when At that time, the k-th modal component is determined to be a noise-dominant component and is removed; thus, an index set of effective modal components is constructed. .
4. The method for extracting and recognizing visual features of UAV radio frequency signals according to claim 1, characterized in that: The radio frequency signal in step four is: 。 5. The method for extracting and recognizing visual features of UAV radio frequency signals according to claim 1, characterized in that: In step five, assume the one-dimensional time series of the RF signal after denoising is as follows: in, For the sequence length; for the sequence Perform normalization and scale it to an interval. The normalized sequence is obtained. ; For normalized sequences Perform polar coordinate transformation to obtain angle sequence ; Normalized sequence With the sequence of inverse cosine values Introducing polar coordinates to construct a polar radius term ; Based on angle sequence Construct the Gram angle field matrix, which consists of the Gram angle sum field and the Gram angle difference field: By using a two-dimensional matrix of Gram angle sum field GASF and Gram angle difference field GADF, a one-dimensional radio frequency signal can be encoded into a two-dimensional image that can effectively characterize the correlation between different time points.
6. The method for extracting and recognizing visual features of UAV radio frequency signals according to claim 5, characterized in that: The formula for calculating the normalized sequence is: in, and Sequences The maximum and minimum values.
7. The method for extracting and recognizing visual features of UAV radio frequency signals according to claim 5, characterized in that: Polar coordinate transformation is implemented using the inverse cosine function, and its calculation formula is as follows: in Indicates sample The value of the inverse cosine.
8. The method for extracting and recognizing visual features of UAV radio frequency signals according to claim 1, characterized in that: The convolutional neural network in step six includes: four convolutional layers. The first convolutional layer has 32 kernels and a kernel size of 3×3. The second, third, and fourth convolutional layers have 64, 128, and 256 kernels, respectively, and each has a convolutional size of 3×3. Each convolutional layer is followed by a pooling layer with a pooling stride of 2. The convolutional outputs are flattened and connected to a fully connected layer with 512 hidden nodes. A Dropout layer with a Dropout ratio of 0.5 is then set after the fully connected layer. Finally, a Softmax classification layer is set to output the probabilities of different drone categories, and the category with the highest probability is used as the drone type recognition result.
9. The method for extracting and recognizing visual features of UAV radio frequency signals according to claim 1, characterized in that: The training process of a convolutional neural network uses the cross-entropy loss function as the optimization objective. The network weight parameters are updated through backpropagation until the loss function converges or reaches the preset number of training rounds, thereby obtaining a trained convolutional neural network classification model.