Lightweight and real-time OcuSync protocol drone identification method based on preamble extraction

Abstract

The embodiments of the present application relate to the field of drone identification technology, and more particularly to a lightweight, real-time OcuSync protocol drone identification method based on preamble code extraction, comprising: collecting drone signals; converting the collected drone signals into time-frequency spectra through STFT, and extracting image transmission signals from the time-frequency spectra using a dual-window sliding threshold judgment algorithm based on the time-frequency spectra; extracting pilot signals from the image transmission signals using margin factor features based on cross-correlation spectra; performing spectral cropping on the pilot signals to obtain enhanced pilot signals; inputting the enhanced pilot signals into a drone model classification model to obtain a preliminary recognition result output by the drone model classification model; and jointly judging multiple preliminary recognition results through voting to obtain a final drone model determination result. This method achieves efficient classification and recognition of different drone models and different individuals of the same model.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of drone identification technology, and more particularly to a lightweight and real-time OcuSync protocol drone identification method based on preamble code extraction. Background Art

[0002] In recent years, drone technology has rapidly developed and has been widely used in military and civilian applications such as surveillance, search and rescue, logistics and transportation, agriculture, and environmental protection. Therefore, the research and development of drone detection and identification technology is crucial. It is a core prerequisite for drone regulation and a key guarantee for ensuring airspace security and preventing potential threats.

[0003] Currently, drone detection and identification technologies can be categorized into four main categories: vision-based, sound-based, radar-based, and radio frequency-based. Vision-based detection and identification is susceptible to changes in lighting conditions and obstructions, sound-based detection and identification is limited by its effective range, and radar-based detection and identification is an active detection method, which is costly and lacks the stealthiness of monitoring. Radio frequency-based detection and identification identifies wireless devices by detecting their radio frequency signals. It has been used to classify various wireless devices, including LoRa, ZigBee, and WiFi. It is suitable for drones that use radio frequency signals for uplink and downlink communications and represents a highly promising drone identification technology.

[0004] Unlike traditional RFID systems, drone classification based on RFID is considered a blind identification task because drone communication protocols are typically confidential. The difficulty in obtaining prior information about drone protocol parameters and extracting bit-level information from drone signals is also difficult. Manually extracting features from RF signals is also challenging, posing significant challenges to the design of drone RFID classification systems.

[0005] The rise of deep learning technology has brought new breakthroughs to RFID-based drone classification. Deep learning is a data-driven approach that doesn't rely on prior knowledge. By directly feeding collected drone signals into neural network training, or by converting the signals into specific representations before feeding them into the neural network for training, it automatically discovers potential, potentially overlooked, features within large amounts of drone data and enables classification.

[0006] For drone detection and identification based on radio frequency identification, many research teams have focused on inputting the transformed form of drone data signals into deep learning models or extracting features from drone data signals and then inputting them into the model for learning, hoping that the model can extract potential features to achieve drone classification. For example, Rubina Akter et al. used a convolutional neural network containing one-dimensional convolutional layers to extract feature maps of different scales from the DroneRF dataset for drone detection and classification. The model achieved an average classification accuracy of 92.5% when processing data containing background noise and three different drone signals. Chaoqun Li et al. combined time-frequency spectrum and bispectral spectrum features and performed feature extraction and feature fusion through a dual-channel CNN. When the signal-to-noise ratio was 30 dB, the method achieved a classification accuracy of 99.47% for a dataset containing 17 types of drone remote controls. Kenneth Bremnes et al. proposed a radio frequency fingerprint recognition method based on fixed-boundary empirical wavelet subbands and used CNN for drone classification. The method was tested on 15 types of drones and achieved an average classification accuracy of 97.25%. Zhenxin Cai et al. designed a lightweight backbone network composed of LMSC (lightweight multi-scale convolution) blocks, thereby proposing a lightweight model for drone radio frequency identification, which to some extent takes into account the deployment problem of drone detection and identification systems.

[0007] The inventors of this application have discovered that the currently proposed drone detection and identification methods based on radio frequency identification have the following problems.

[0008] First, there's poor versatility. Currently proposed methods are primarily designed for specific datasets, making their high performance difficult to generalize to a specific type of drone or manufacturer. In particular, there's insufficient research on drone identification, which utilizes the advanced proprietary OcuSync protocol, a popular market share. This hinders practical application of these high-performance methods.

[0009] Second, the integration of expert knowledge and deep learning is insufficient. Existing research still has room for improvement in effectively combining the strengths of expert knowledge and deep learning. Existing research has primarily focused on inputting transformed drone data signals into deep learning models or extracting features from drone data signals before feeding them into the model for learning. The hope is that the model can extract latent features to enable drone classification. However, in addition to designing signal transformation and feature extraction, expert knowledge should also play a greater role in selecting the data to be fed into deep learning models for training. How to incorporate expert knowledge into data selection and model design to improve model performance and robustness is a pressing issue.

[0010] Third, the importance of pilot signals has been overlooked, leading to a distorted understanding of the mechanisms underlying individual identification. Previous studies have mistakenly attributed the distinction between different drones of the same model to RFF (radio frequency fingerprint), believing that RFF is the key to distinguishing different individuals of the same model. However, the preamble, or pilot symbol, within a drone's image transmission signal is the key to identifying a drone. This is especially true under the OcuSync protocol, which exhibits multiple modes, and signals from different individuals of the same model are inherently distinct. Previous studies have overlooked the importance of pilot signals and the multi-modal nature of the OcuSync protocol, leading to a distorted understanding of the mechanisms underlying individual identification.

[0011] Fourth, system deployment and efficiency issues have not been fully considered. Existing research has given little consideration to the deployment of drone detection and identification systems and the efficiency of system detection and identification. Although Zhenxin Cai et al. proposed a lightweight model, the lightweight model is only a small part of the deployment of drone detection and identification systems. More critical is the need to consider hardware-specific stream processing solutions. Furthermore, performing a short-time Fourier transform on drone signals and then feeding them into a convolutional neural network for training and identification is considered the optimal solution for matching signal representation with deep learning models. However, the amount of signal data required for a single detection and identification is enormous, resulting in low detection and identification efficiency and difficulty meeting real-time requirements. Summary of the Invention

[0012] In view of this, the embodiments of the present application propose a lightweight and real-time OcuSync protocol drone identification method based on preamble code extraction, aiming to provide a high-efficiency, low-cost, and universal identification method suitable for OcuSync protocol drones. This method extracts and identifies pilot symbols in drone image transmission signals, thereby achieving efficient classification and identification of different models and different individuals of the same model.

[0013] In the first aspect, an embodiment of the present application proposes a lightweight and real-time OcuSync protocol drone identification method based on preamble code extraction, the method comprising: collecting drone signals through a radio frequency receiving device; converting the collected drone signals into a time-frequency spectrum through STFT (Short Time Fourier Transform), and extracting the image transmission signal from the time-frequency spectrum using a double-window sliding threshold judgment algorithm based on the time-frequency spectrum; using the margin factor feature based on the cross-correlation spectrum to distinguish the pilot signal and the non-pilot signal in the image transmission signal, thereby extracting the pilot signal from the image transmission signal; performing spectral cropping on the pilot signal to obtain an enhanced pilot signal; inputting the enhanced pilot signal into a pre-trained drone model classification model to obtain a preliminary recognition result output by the drone model classification model; and jointly judging multiple preliminary recognition results by voting to obtain the final drone model determination result.

[0014] Optionally, the collected drone signal is converted into a time-frequency spectrum through STFT, and a dual-window sliding threshold judgment algorithm based on the time-frequency spectrum is used to extract the image transmission signal from the time-frequency spectrum, including:

[0015] Initialize the first intermediate variable η left , the second intermediate variable η mid and the third intermediate variable η right , and preset the signal change trend judgment threshold η threld , the time length threshold T of the image transmission signal threld , and bandwidth threshold B threld ;

[0016] Initialize the variable T at the start time of the recording signal start , variable T at the end time of the recording signal end , variable F that records the starting position of the signal on the spectrum axis start , and the variable F that records the end position of the signal on the spectrum axis end ;

[0017] The collected drone signal of length FFTsize is converted into a time-frequency spectrum through STFT. The spectrum maximum of the time-frequency spectrum is extracted and sent to a sliding window of length L. If the sliding window is not full, the spectrum maximum is added to the empty space closest to the leftmost side of the sliding window. If the sliding window is full, the value on the leftmost side of the sliding window is removed and the new maximum value is added to the rightmost side of the sliding window.

[0018] Whenever the sliding window is full, the energy value E of the left half of the sliding window is calculated left and the energy value E of the right half right, and calculate the energy ratio η, η=E right / E left , η mid Update to η left , η right Update to η mid , update η to η right ;

[0019] According to η lefr ,η mid and η righr The relationship between the size of η is used to judge the changing trend of η. If η right >η mid >η left , then it is determined that η continues to increase. If η right <η mid <η left , then it is determined that η continues to decrease;

[0020] When T start =0, determine that η continues to increase and η>η threld When the signal starts, find the starting time point T start , start accumulating spectrum data, when it is determined that η continues to decrease and η<1 / η threld When the signal ends, find the end time point T end , stop accumulating spectrum data;

[0021] Using the spectrum data sequence, repeat the sliding window operation to find the starting point F of the signal on the spectrum axis. start and end point F end , and according to T start 、T end 、F start and F end , calculate the time length T of the signal x and bandwidth B x ;

[0022] If R x >T threld And B x >B threld , it is confirmed that the image transmission signal is detected, and according to T start and T end , extract the image transmission signal from the time-frequency spectrum.

[0023] Optionally, using a margin factor feature based on a cross-correlation spectrum to distinguish a pilot signal from a non-pilot signal in the image transmission signal, thereby extracting the pilot signal from the image transmission signal, includes:

[0024] Consider each line of the image transmission signal as a time segment of the smallest unit signal, and calculate the normalized cross-correlation function between two adjacent time segments. Suppose the total number of time segments is M, and the time segment s is i With the previous time segment s i-1 The normalized cross-correlation function between i (τ), R i (τ) is calculated by the following formula:

[0025]

[0026] Where τ represents the time delay, μ i Indicates s i The mean value, μ i-1 Indicates s i-1 The mean of i Indicates s i The standard deviation, σ i-1 Indicates s i-1 The standard deviation of , E(·) means to obtain the expectation;

[0027] The normalized cross-correlation function corresponding to each time segment is stored in a cross-correlation matrix R matrix middle;

[0028] Calculate the margin factor of the normalized cross-correlation function corresponding to each time segment respectively, and set the time segment s i The corresponding margin factor is MF i , MF i Calculated by the following formula:

[0029]

[0030] Among them, max(·) means to find the maximum value, LE means R i the length of (τ);

[0031] The margin factors corresponding to each time segment are combined into a feature vector MF;

[0032] The clustering algorithm is used to classify MF, and all margin factors are divided into two categories to obtain the classification labels corresponding to each time segment. The two categories represent high correlation areas and low correlation areas, respectively. The clustering goal is to minimize the sum of squared errors within the class. The clustering goal is expressed by the formula:

[0033]

[0034] Among them, G represents the target of clustering, c k represents the cluster center of the kth category, I(·) is the indicator function, if MF i Belongs to the kth category, then I(MFi ∈c k )=1, if MF i does not belong to the kth category, then I(MF i ∈c k )=0;

[0035] Based on the margin factor, pulse detection is performed on each time segment to obtain the pulse detection result corresponding to each time segment. Let the time segment s i The corresponding pulse detection result is P i , P i Calculated by the following formula:

[0036]

[0037] Among them, max(c1,c2)-1 represents the preset clustering decision threshold, P i =1 means time segment s i The corresponding signal is the pilot signal, P i =0 means time segment s i The corresponding signal is a non-pilot signal;

[0038] The pulse detection results corresponding to each time segment are combined into a pulse sequence P, where P i =1 corresponds to the position index corresponding to the time point of the pilot signal in the image transmission signal;

[0039] Starting from the first pilot signal detected, the time-frequency spectra of N consecutive pilot signals are combined to obtain a pilot symbol;

[0040] Each pilot symbol is filtered, and pilot symbols whose actual occupied number of rows is not equal to N are removed, and finally the pilot signal is extracted from the image transmission signal.

[0041] Optionally, performing spectrum clipping on the pilot signal to obtain an enhanced pilot signal includes:

[0042] The pilot signal is converted to the frequency domain using DFT (Discrete Fourier Transform) to obtain the spectrum of the pilot signal. The spectrum of the pilot signal is then preprocessed, including spectrum smoothing and spectrum binarization, to obtain a preprocessed spectrum.

[0043] Detect the widest pulse region in the pre-processed spectrum and determine the starting frequency point f of the widest pulse region start and the end frequency point f end , and calculate the center frequency point f center , f center =(f start +f end ) / 2;

[0044] f center As the center, the spectrum of the pilot signal is trimmed according to the preset retention width to obtain the trimmed spectrum R cropped (f), R cropped (f) is expressed by the formula:

[0045] R cropped (f)=R(f)for f∈[f center -Δf,f center +Δf];

[0046] Wherein, R(f) represents the frequency spectrum of the pilot signal, and Δf represents the preset retention width;

[0047] R cropped (f) Convert to the time domain through IDFT to obtain the enhanced pilot signal.

[0048] Optionally, the network structure of the pre-trained drone model classification model is UAVResNet improved on the basis of ResNet, which is specifically composed of a Layer0, a Layer1, an adaptive average pooling layer and an output layer; Layer0 is composed of a 3×3 convolution layer and a 3×3 maximum pooling layer, and the 3×3 convolution layer is followed by a BN normalization layer and a ReLU activation function layer; Layer1 is composed of a Bottleneck structure with an output channel number of 64, and the Bottleneck structure is specifically composed of two 1×1 convolution layers, a 3×3 convolution layer and a residual connection; the adaptive average pooling layer is used to resize the feature map output by Layer1 to 1×1; the output layer is composed of a fully connected layer and a Softmax layer, which is used to perform classification based on the output of the adaptive average pooling layer, and finally output the classification probability.

[0049] Optionally, the pre-trained UAV model classification model is trained by the following steps: collecting a large number of UAV pilot signals, performing spectral cropping on the collected UAV pilot signals, and obtaining enhanced UAV pilot signals; using the enhanced UAV pilot signals as training samples, and marking them with true labels to obtain a training sample set; constructing UAVResNet as the initial model; inputting the training samples in the training sample set into the initial model, iteratively training the initial model until convergence, and obtaining a trained UAV model classification model.

[0050] Optionally, a joint judgment is made on multiple preliminary identification results through voting to obtain a final drone model determination result, including: determining the confidence of each preliminary identification result based on the signal-to-noise ratio of the pilot signal corresponding to each preliminary identification result, the higher the signal-to-noise ratio of the pilot signal corresponding to the preliminary identification result, the higher the confidence of the preliminary identification result; and performing weighted voting based on the confidence of each preliminary identification result to obtain the final drone model determination result.

[0051] The embodiments of the present application propose a lightweight and real-time OcuSync protocol drone identification method based on preamble code extraction, which has at least the following advantages compared with the currently proposed drone detection and identification methods.

[0052] First, it has higher recognition efficiency. This application only requires a single pilot symbol to classify and identify drones. However, existing RFID-based drone detection and identification methods require processing the entire time-frequency spectrum (usually containing multiple pilot symbols and other signal data) for identification, which results in a huge amount of data. Therefore, the recognition efficiency of this application has been significantly improved, enabling faster detection and identification.

[0053] Second, it has lower deployment costs. The data input to the drone model classification model in this application is only a single pilot symbol, which places low requirements on the model. The number of model parameters and the amount of computation are relatively small, effectively reducing deployment costs and making this method easier to deploy on resource-constrained hardware platforms.

[0054] Third, better versatility. This application is designed for OcuSync protocol drones. By extracting pilot symbols for identification, it can be applied to different models under the protocol and different individuals of the same model, with good versatility.

[0055] Fourth, it more fully incorporates expert knowledge. This application fully utilizes expert knowledge and selects pilot symbols as the input of the drone model classification model. This choice is based on a deep understanding of drone communication protocols, reflects the important role of expert knowledge, and improves the accuracy and efficiency of recognition.

[0056] In the second aspect, an embodiment of the present application proposes a lightweight and real-time OcuSync protocol drone identification system based on preamble code extraction, the system including: a radio frequency receiving device for collecting drone signals; an image transmission signal extraction module for converting the collected drone signals into a time-frequency spectrum through STFT, and extracting the image transmission signal from the time-frequency spectrum using a double-window sliding threshold judgment algorithm based on the time-frequency spectrum; a pilot signal extraction module for distinguishing the pilot signal and non-pilot signal in the image transmission signal using a margin factor feature based on the cross-correlation spectrum, thereby extracting the pilot signal from the image transmission signal; a data enhancement module for performing spectrum cropping on the pilot signal to obtain an enhanced pilot signal; a preliminary identification module for inputting the enhanced pilot signal into a pre-trained drone model classification model to obtain a preliminary identification result output by the drone model classification model; a joint judgment module for jointly judging multiple preliminary recognition results by voting to obtain a final drone model judgment result.

[0057] In a third aspect, an embodiment of the present application proposes an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute a lightweight and real-time OcuSync protocol drone identification method based on preamble code extraction as described in the first aspect above.

[0058] In a fourth aspect, an embodiment of the present application proposes a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it can implement a lightweight and real-time OcuSync protocol drone identification method based on preamble code extraction as described in the first aspect above.

[0059] It can be understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the related technologies, the following is a brief introduction to the drawings required for use in the embodiments of the present application or the description of the related technologies. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0061] Figure 1 This is a flowchart of a lightweight and real-time OcuSync protocol drone identification method based on preamble extraction provided in one embodiment of the present application;

[0062] Figure 2 is a schematic diagram of extracting image transmission signals provided in one embodiment of the present application;

[0063] Figure 3 1 is a schematic diagram comparing the cross-correlation function between pilot signals and the cross-correlation function between non-pilot signals provided in one embodiment of the present application;

[0064] Figure 4 is a schematic diagram of extracting a pilot signal provided in one embodiment of the present application;

[0065] Figure 5 is a schematic diagram of achieving data enhancement through spectrum clipping provided in one embodiment of the present application;

[0066] Figure 6 is a schematic diagram of the network structure of UAVResNet provided in one embodiment of the present application;

[0067] Figure 7 1 is a schematic diagram of the structure of the bottleneck in the UAVResNet provided in one embodiment of the present application;

[0068] Figure 8 : This is a schematic diagram comparing the average drone recognition accuracy under different FFT points and different data enhancement methods provided in an embodiment of the present application;

[0069] Figure 9 This is a comparative diagram of the average drone recognition accuracy of different filter operations and joint decisions under different SNRs provided in one embodiment of the present application;

[0070] Figure 10 This is a structural diagram of a lightweight and real-time OcuSync protocol drone identification system based on preamble extraction provided in another embodiment of the present application;

[0071] Figure 11 It is a structural diagram of an electronic device provided in another embodiment of the present application. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the embodiments of the present application will be described in detail below with reference to the accompanying drawings. In the various embodiments of the present application, many technical details are proposed to enable the reader to better understand the present application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solutions claimed in the present application can be implemented. The division of the following embodiments is only for the convenience of description and should not constitute any limitation on the specific implementation of the present application. The various embodiments can be combined with each other and referenced to each other under the premise of no contradiction.

[0073] An embodiment of the present application proposes a lightweight and real-time OcuSync protocol drone identification method based on preamble code extraction, which is applied to an electronic device, wherein the electronic device can be a terminal or a server. In this embodiment and the following embodiments, the electronic device is described using the server as an example. The implementation details of the lightweight and real-time OcuSync protocol drone identification method based on preamble code extraction proposed in this embodiment are specifically described below. The following content is only the implementation details provided for the convenience of understanding and is not necessary for the implementation of this solution.

[0074] The specific process of the lightweight and real-time OcuSync protocol drone identification method based on preamble extraction proposed in this embodiment can be as follows: Figure 1 Shown, including:

[0075] Step 101: Collect drone signals through a radio frequency receiving device.

[0076] In the specific implementation, the server first needs to collect drone signals through a radio frequency receiving device. The radio frequency receiving device used can be USRP (Universal Software Radio Peripheral) and the like.

[0077] In step 102, the collected drone signal is converted into a time-frequency spectrum through STFT, and a double-window sliding threshold judgment algorithm based on the time-frequency spectrum is used to extract the image transmission signal from the time-frequency spectrum.

[0078] In the specific implementation, after collecting the drone signal, the server needs to convert the collected drone signal (time domain IQ signal) into a time-frequency spectrum through STFT, and use the double-window sliding threshold judgment algorithm based on the time-frequency spectrum to extract the image transmission signal from the time-frequency spectrum.

[0079] In one example, when the server extracts the image transmission signal, it first needs to initialize the first intermediate variable η left , the second intermediate variable η mid and the third intermediate variable η right, and preset the signal change trend judgment threshold η threld , the time length threshold T of the image transmission signal threld , and bandwidth threshold B threld At the same time, it is also necessary to initialize the variable T at the starting time of the recording signal start , variable T at the end time of the recording signal end , variable F that records the starting position of the signal on the spectrum axis start , and the variable F that records the end position of the signal on the spectrum axis end Among them, T start 、T end 、F start and F end The initial value of η is 0. left ,η mid and η right The initial value of η is 1. left ,η mid and η right The energy ratios of the last three cycles are recorded separately to make trend judgments.

[0080] After initialization is complete, the server needs to convert the collected drone signal of length FFTsize into a time-frequency spectrum through STFT, extract the spectrum maximum of the time-frequency spectrum, and send the spectrum maximum to a sliding window of length L. If the sliding window is not full, the spectrum maximum is added to the empty space closest to the leftmost position in the sliding window. If the sliding window is full, the leftmost value of the sliding window is removed and the new maximum value is added to the rightmost position of the sliding window.

[0081] Whenever the sliding window is full, the server needs to calculate the energy value E of the left half of the sliding window separately left and the energy value E of the right half right , and calculate the energy ratio η, η=E right / E left , then update the three intermediate variables, and set η mid Update to η left , η right Update to η mid , update η to η right .

[0082] Next, the server needs to left ,η nid and η right The relationship between the size of η and the change trend of η can be used to accurately identify the starting time point and the ending time point of the signal. right >η mid >η left , then it is determined that η continues to increase, indicating that the signal strength may be increasing.right <η mid <η left , it is determined that η is continuously decreasing, indicating that the signal strength may be decreasing. In other cases, it means that η does not change significantly and the trend of signal strength change is not clear.

[0083] When the server detects T start =0, determine that η continues to increase and η>η threld When the signal is found, the starting time point T is determined. start , then begins to accumulate spectrum data. When the server determines that η continues to decrease and η<1 / η threld When the signal end time T is found, end , at this time, it is necessary to stop accumulating spectrum data.

[0084] Subsequently, the server uses the spectrum data sequence to repeat the above sliding window operation to find the starting point F of the signal on the spectrum axis. start and end point F end , and according to T start 、T end 、F start and F end , calculate the time length T of the signal x and bandwidth B x .

[0085] Finally, the server is based on T x 、B x 、T threld and B threld , to determine whether the image transmission signal is detected. If T x >T threld And B x >B threld , the server confirms that the image transmission signal is detected and start and T end , extract the image transmission signal from the time-frequency spectrum.

[0086] In summary, the server achieves robust, accurate, and adaptive recognition of image transmission signals through dual-window sliding and trend-based threshold determination, effectively reducing false detection and missed detection rates and improving the reliability of the image transmission signal extraction process. The algorithm's threshold is pre-set and does not require dynamic adjustment, but the trend determination results influence the algorithm's application of the threshold, thereby indirectly achieving adaptability. Figure 2 Schematic diagram of the process of extracting the image transmission signal of the DJPhantom4ProV2 drone.

[0087] Step 103 : Using the margin factor feature based on the cross-correlation spectrum to distinguish the pilot signal from the non-pilot signal in the image transmission signal, thereby extracting the pilot signal from the image transmission signal.

[0088] In a specific implementation, after extracting the image transmission signal, the server needs to use the margin factor feature based on the cross-correlation spectrum to distinguish the pilot signal and non-pilot signal in the image transmission signal, thereby extracting the pilot signal from the image transmission signal.

[0089] In one example, the server regards each line of the image transmission signal as a time segment of the minimum unit signal, and calculates the normalized cross-correlation function between two adjacent time segments. Let the total number of time segments be M, and the time segment s be i With the previous time segment s i-1 The normalized cross-correlation function between i (τ), R i (τ) is calculated by the following formula:

[0090]

[0091] Where τ represents the time delay, μ i Indicates s i The mean value, μ i-1 Indicates s i-1 The mean of i Indicates s i The standard deviation, σ i-1 Indicates s i-1 The standard deviation can be used to estimate the energy of the signal very well, and E(·) represents the expectation.

[0092] The comparison of the cross-correlation function between the pilot signal and the non-pilot signal can be shown as follows: Figure 3 As shown, when s i and s i-1 When both are pilot signals, R i (τ) shows an obvious multi-peak shape. i and s i-1 When not all pilot signals are used, R i The peak of (τ) is not obvious.

[0093] In order to facilitate processing and analysis, the server stores the normalized cross-correlation function corresponding to each time segment in a cross-correlation matrix R matrix In, R matrix It is obtained by the time-frequency spectrum of the signal (the signal is STFT), R matrix The first row of the time-frequency spectrum is obtained by cross-correlating the second row (corresponding to the second time slice) with the first row, and the final R matrix There will be one less line than the time-frequency spectrum.

[0094] In order to quantify the multi-peak characteristics of the cross-correlation function, the server introduces a feature called margin factor to reflect the prominence of the cross-correlation function peak. The server calculates the margin factor of the normalized cross-correlation function corresponding to each time segment. Let the time segment s i The corresponding margin factor is MF i , MF i Calculated by the following formula:

[0095]

[0096] Among them, max(·) means to find the maximum value, LE means R i The length of (τ).

[0097] Similarly, in order to facilitate processing and analysis, the server combines the margin factors corresponding to each time segment into a feature vector MF. MF=[MF1, MF2,…, MF M ]. This feature vector will be used in subsequent cluster analysis to distinguish pilot signals from non-pilot signals.

[0098] Next, the server uses a clustering algorithm (such as the K-Means clustering algorithm) to classify the MF, dividing all margin factors into two categories (if the K-Means clustering algorithm is used, K must be set to 2). The classification labels corresponding to each time segment are obtained. The two categories represent high-correlation areas and low-correlation areas, respectively. The clustering goal is to minimize the sum of squared errors within the class. The clustering goal is expressed as follows:

[0099]

[0100] Among them, G represents the clustering target, c k represents the cluster center of the kth category, I(·) is the indicator function, if MF i Belongs to the kth category, then I(MF i ∈c k )=1, if MF i does not belong to the kth category, then I(MF i ∈c k )=0.

[0101] Through cluster analysis, the server obtains the classification label corresponding to each time segment. The server continues to perform pulse detection on each time segment based on the margin factor and obtains the pulse detection result corresponding to each time segment. Let the time segment s i The corresponding pulse detection result is P i , P i Calculated by the following formula:

[0102]

[0103] Among them, max(c1,c2)-1 represents the preset clustering decision threshold, P i =1 means time segment s i The corresponding signal is the pilot signal, P i =0 means time segment s i The corresponding signal is a non-pilot signal.

[0104] After completing the pulse detection, the server needs to locate the pilot signal and combine the pilot symbols. The server combines the pulse detection results corresponding to each time segment into a pulse sequence P, where P i = 1 corresponds to the time point number of the corresponding pilot signal in the image transmission signal. When FFTsize is fixed and equal to 0, the number of rows occupied by a pilot symbol in the time-frequency spectrum is also fixed at N. Therefore, after detecting the position of the pilot signal in the time-frequency spectrum, the server needs to observe the number of rows occupied by each pilot symbol in the time-frequency spectrum. Then, starting from the first detected pilot signal, the time-frequency spectrum of N consecutive pilot signals is combined to obtain a pilot symbol.

[0105] Finally, the server filters each pilot symbol (also known as the filter operation), removes the pilot symbols whose actual number of occupied rows is not equal to N, and finally extracts the pilot signal from the image transmission signal.

[0106] In summary, the server accurately extracts the pilot signal from the time-frequency spectrum of the image transmission signal, providing a reliable data basis for subsequent drone identification. The pilot signal extraction process can be as follows: Figure 4 shown.

[0107] Step 104: Perform spectrum clipping on the pilot signal to obtain an enhanced pilot signal.

[0108] In specific implementations, in wireless communication systems, data augmentation technology generates data for other scenarios from limited data collected in real-world scenarios, thereby improving the model's robustness to signals in highly variable real-world environments. Alternatively, data is preprocessed before being sent to the network, with the expectation that the model can extract higher-quality, more essential features from the processed data, thereby improving the model's robustness. Therefore, after extracting the pilot signal from the image transmission signal, the server needs to perform spectral cropping on the pilot signal to achieve data augmentation and obtain an enhanced pilot signal.

[0109] In one example, data augmentation methods also include frequency offset estimation and compensation, frequency shifting, etc. However, in the drone identification scenario, spectrum cropping performs best. The details of spectrum cropping are described below.

[0110] First, the pilot signal is converted to the frequency domain through DFT to obtain the spectrum of the pilot signal. Then, the spectrum of the pilot signal is preprocessed including spectrum smoothing and spectrum binarization to obtain the preprocessed spectrum.

[0111] Assume r(t) represents the pilot signal, the spectrum of the pilot signal can be represented by R(f), and the DFT conversion process can be expressed by the formula:

[0112]

[0113] Where N is the number of signal sampling points, t is the time index, and f is the frequency index.

[0114] The spectrum smoothing operation is achieved by convolution with the Hanning window w(k). The spectrum smoothing operation is expressed by the formula:

[0115]

[0116] Where γ represents the half-width of the Hanning window, and Smoothed_R(f) represents the smoothed spectrum.

[0117] The spectrum binarization operation is based on the dynamic threshold T Δth accomplish.

[0118] T Δth ={max[Smoothed_R(f)]-min[Smoothed_R(f)]} / 2;

[0119] The spectrum binarization operation is expressed by the formula:

[0120]

[0121] Wherein, Binary_R(f) represents the spectrum after the spectrum binarization operation.

[0122] Next, the server needs to detect the widest pulse area in the preprocessed spectrum and determine the starting frequency point f of the widest pulse area. start and the end frequency point f end , and calculate the center frequency point f center , f center =(f start +f end ) / 2. With f center As the center, the spectrum of the pilot signal is trimmed according to the preset retention width to obtain the trimmed spectrum R cropped (f). R cropped (f) is expressed by the formula:

[0123] R cropped (f)=R(f)forf∈[fcenter -Δf,f center +Δf];

[0124] Wherein, R(f) represents the frequency spectrum of the pilot signal, and Δf represents the preset retention width.

[0125] Finally, the server will R cropped (f) Convert to the time domain through IDFT to obtain the enhanced pilot signal.

[0126] Generally speaking, the bandwidth of drone signals is 10MHz, so Δf can be set to 11MHz. The process of data enhancement by spectrum clipping can be as follows: Figure 5 shown.

[0127] Step 105: Input the enhanced pilot signal into the pre-trained UAV model classification model to obtain a preliminary recognition result output by the UAV model classification model.

[0128] In a specific implementation, after the server obtains the enhanced pilot signal, it can input the enhanced pilot signal into the pre-trained drone model classification model to obtain the preliminary recognition result output by the drone model classification model (preliminary drone model classification result).

[0129] In one example, the network structure of the pre-trained drone model classification model is UAVResNet, which is improved on the basis of ResNet. ResNet is a mature and effective CNN architecture, which is widely used in image recognition tasks. The signal processed by this application is a drone signal. The essential difference between drone signals and images lies in the number of channels and the size of the data. The number of channels of an image is usually 3, and the input size is usually a square size. The number of channels of the time-frequency spectrum of a drone signal is 1, and the input size is a rectangular size with a significant difference in length and width. Therefore, this application proposes a deep convolutional neural network UAVResNet based on the improvement of ResNet, which is specially designed for the characteristics of drone signals and simplifies the network structure. The network structure of UAVResNet can be as follows Figure 6 shown.

[0130] The ResNet network structure usually consists of a Layer0, a Layer1, a Layer2, a Layer3, a Layer4, an adaptive average pooling layer, and an output layer. Layer0 consists of a 3×3 convolution layer and a 3×3 maximum pooling layer. The 3×3 convolution layer is followed by a BN normalization layer and a ReLU activation function layer. Layer1 to Layer4 stack multiple Bottleneck structures. Bottleneck structures are as follows: Figure 7As shown in the figure, each Bottleneck structure contains three convolutional layers (1×1 convolutional layer, 3×3 convolutional layer, 1×1 convolutional layer) and residual connections, with the number of output channels being 64, 128, 256, and 512 respectively. The deep structure of ResNet gives it strong feature expression capabilities, but also brings high computational complexity and memory usage.

[0131] UAVResNet is optimized for the resource constraints of drone platforms and consists of a Layer 0, a Layer 1, an adaptive average pooling layer, and an output layer. Layer 0 consists of a 3×3 convolutional layer and a 3×3 max pooling layer, followed by a batch normalization layer and a ReLU activation function layer. Layer 1 consists of a Bottleneck structure with 64 output channels, consisting of two 1×1 convolutional layers, a 3×3 convolutional layer, and a residual connection. The adaptive average pooling layer is used to resize the feature map output by Layer 1 to 1×1. The output layer consists of a fully connected layer and a softmax layer, which is used to perform classification based on the output of the adaptive average pooling layer and ultimately output the classification probability.

[0132] The key improvement of UAVResNet lies in its streamlined network structure. Compared to ResNet, UAVResNet reduces the number of network layers and bottlenecks, lowering model complexity and resource usage, making it more suitable for drone identification systems with limited computing and storage resources. This targeted optimization maximizes recognition performance while improving its deployment feasibility in drone identification systems.

[0133] In one example, a pre-trained drone model classification model is trained through the following steps.

[0134] First, the server collects a large number of drone pilot signals and performs spectrum cropping on them to generate enhanced drone pilot signals. These enhanced drone pilot signals are then used as training samples and annotated with true labels to create a training sample set. Next, the server constructs UAVResNet as the initial model. Finally, the server inputs the training samples from the training sample set into the initial model and iteratively trains the initial model until convergence, resulting in a fully trained drone model classification model.

[0135] Step 106 , a plurality of preliminary identification results are jointly judged by voting to obtain a final drone model determination result.

[0136] In practice, to improve the accuracy and reliability of drone identification, this application introduces a joint decision mechanism. This mechanism comprehensively analyzes multiple pilot signals within a complete drone signal, then uses pre-trained drone model classification models to identify each, obtaining multiple preliminary recognition results. These preliminary results are then combined through voting to arrive at the final drone model determination. The core idea of ​​the joint decision mechanism is to leverage information fusion to improve recognition accuracy and robustness.

[0137] In one example, multiple preliminary recognition results are jointly judged by voting to obtain the final drone model determination result, that is, the model that appears most frequently in the multiple preliminary recognition results is taken as the final result.

[0138] In one example, the server can determine the confidence level of each preliminary identification result based on the signal-to-noise ratio of the pilot signal corresponding to each preliminary identification result. The higher the signal-to-noise ratio of the pilot signal corresponding to the preliminary identification result, the higher the confidence level of the preliminary identification result. A weighted vote is then performed based on the confidence level of each preliminary identification result to determine the final drone model, resulting in a more reliable judgment. By comprehensively considering the judgment results of multiple pilot signals in a complete image transmission signal, the risk of misjudgment due to signal interference, distortion, or recognition uncertainty is effectively reduced.

[0139] This embodiment proposes a lightweight and real-time OcuSync protocol drone identification method based on preamble code extraction, which has at least the following advantages compared with the currently proposed drone detection and identification methods.

[0140] First, it has higher recognition efficiency. This application only requires a single pilot symbol to classify and identify drones. However, existing RFID-based drone detection and identification methods require processing the entire time-frequency spectrum (usually containing multiple pilot symbols and other signal data) for identification, which results in a huge amount of data. Therefore, the recognition efficiency of this application has been significantly improved, enabling faster detection and identification.

[0141] Second, it has lower deployment costs. The data input to the drone model classification model in this application is only a single pilot symbol, which places low requirements on the model. The number of model parameters and the amount of computation are relatively small, effectively reducing deployment costs and making this method easier to deploy on resource-constrained hardware platforms.

[0142] Third, better versatility. This application is designed for OcuSync protocol drones. By extracting pilot symbols for identification, it can be applied to different models under the protocol and different individuals of the same model, with good versatility.

[0143] Fourth, it more fully incorporates expert knowledge. This application fully utilizes expert knowledge and selects pilot symbols as the input of the drone model classification model. This choice is based on a deep understanding of drone communication protocols, reflects the important role of expert knowledge, and improves the accuracy and efficiency of recognition.

[0144] The step division of the above various methods is only for the purpose of clear description. During implementation, they can be combined into one step or some steps can be split and decomposed into multiple steps. As long as they include the same logical relationship, they are all within the scope of protection of this application; adding insignificant modifications or introducing insignificant designs to the algorithm or process without changing the core design of the algorithm and process are all within the scope of protection of this application.

[0145] The value of FFTsize has an impact on the performance of the algorithm. In order to study the impact of the value of FFTsize on the performance of the drone identification algorithm proposed in this application, we systematically changed the value of FFTsize under a benchmark experimental condition and recorded the performance indicators of the algorithm under different FFTsizes. The conditions of the benchmark experiment were set as follows: the signal-to-noise ratio was 15dB, no data enhancement method was used, and the filter operation and joint decision mechanism were not enabled. The values ​​of FFTsize were 1024, 2048 and 4096 respectively. The experimental results are shown in the figure. Figure 8 As shown. Figure 8 It can be seen from the figure that the drone recognition algorithm proposed in this application has stable performance.

[0146] Next, we conducted a combination experiment of Filter operation and joint decision mechanism. Different combinations produced different results under different signal-to-noise ratio environments. The experimental results are as follows: Figure 9 As shown. Figure 9 It can be seen from the figure that the drone recognition algorithm proposed in this application has the best performance.

[0147] Another embodiment of the present application proposes a lightweight and real-time OcuSync protocol drone identification system based on preamble code extraction. The implementation details of the lightweight and real-time OcuSync protocol drone identification system based on preamble code extraction proposed in this embodiment are described in detail below. The following content is only the implementation details provided for the convenience of understanding and is not necessary for the implementation of this example.

[0148] Figure 10 2 is a schematic diagram of the structure of a lightweight, real-time OcuSync protocol drone identification system based on preamble code extraction proposed in this embodiment. The system includes: a radio frequency receiving device 201, an image transmission signal extraction module 202, a pilot signal extraction module 203, a data enhancement module 204, a preliminary identification module 205, and a joint decision module 206.

[0149] The radio frequency receiving device 201 is used to collect drone signals.

[0150] The image transmission signal extraction module 202 is used to convert the collected drone signal into a time-frequency spectrum through STFT, and use a double-window sliding threshold judgment algorithm based on the time-frequency spectrum to extract the image transmission signal from the time-frequency spectrum.

[0151] The pilot signal extraction module 203 is configured to distinguish the pilot signal from the non-pilot signal in the image transmission signal by using the margin factor feature based on the cross-correlation spectrum, thereby extracting the pilot signal from the image transmission signal.

[0152] The data enhancement module 204 is configured to perform spectrum cropping on the pilot signal to obtain an enhanced pilot signal.

[0153] The preliminary recognition module 205 is used to input the enhanced pilot signal into the pre-trained drone model classification model to obtain a preliminary recognition result output by the drone model classification model.

[0154] The joint decision module 206 is used to make a joint decision on the multiple preliminary recognition results by voting to obtain a final UAV model decision result.

[0155] It is worth mentioning that all modules involved in this embodiment are logical modules. In actual applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, to highlight the innovation of this application, this embodiment does not include units that are not closely related to solving the technical problem proposed by this application. However, this does not mean that other units do not exist in this embodiment.

[0156] It is not difficult to find that this embodiment is a system embodiment corresponding to the above-mentioned method embodiment, and this embodiment can be implemented in conjunction with the above-mentioned method embodiment. The relevant technical details and technical effects mentioned in the above-mentioned embodiments are still valid in this embodiment, and to reduce repetition, they are not repeated here. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above-mentioned embodiments.

[0157] Another embodiment of the present application provides an electronic device, the specific structure of which is as follows: Figure 11 As shown, it includes: at least one processor 301; and a memory 302 communicatively connected to the at least one processor 301; wherein the memory 302 stores instructions that can be executed by the at least one processor 301, and the instructions are executed by the at least one processor 301 to enable the at least one processor 301 to execute a lightweight and real-time OcuSync protocol drone identification method based on preamble code extraction as described in the above method embodiments.

[0158] The memory and processor can be connected using a bus, which can include any number of interconnected buses and bridges. The bus connects various circuits of one or more processors and memories. The bus can also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits. These are all well known in the art and will not be described further in this article. The bus interface is responsible for providing an interface between the bus and the transceiver. The transceiver can be a single component or multiple components, such as multiple receivers and transmitters, providing a unit for communicating with various other devices on a transmission medium. Data processed by the processor is transmitted over a wireless medium via an antenna. Furthermore, the antenna also receives data and transmits it to the processor.

[0159] The processor is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory can be used to store data used by the processor when performing operations.

[0160] Another embodiment of the present application proposes a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it can implement a lightweight and real-time OcuSync protocol drone identification method based on preamble code extraction as described in the above method embodiments.

[0161] That is, those skilled in the art will understand that all or part of the steps in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a program, which is stored in a storage medium and includes a number of instructions for causing a device (such as a single-chip microcomputer, chip, etc.) or a processor to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a USB flash drive, a mobile hard drive, a ROM (Read-Only Memory), a RAM (Random Access Memory), a magnetic disk, or an optical disk, etc., various media that can store program code.

[0162] Those skilled in the art will appreciate that the above embodiments are specific embodiments for implementing the present application, and that in actual applications, various changes may be made thereto in form and detail without departing from the spirit and scope of the present application.

Claims

1. A lightweight and real-time OcuSync protocol drone identification method based on preamble extraction, characterized by: The method comprises: Collect drone signals through radio frequency receiving equipment; The collected drone signal is converted into a time-frequency spectrum through STFT, and the image transmission signal is extracted from the time-frequency spectrum using a double-window sliding threshold judgment algorithm based on the time-frequency spectrum. The pilot signal and non-pilot signal in the image transmission signal are distinguished by using the margin factor feature based on the cross-correlation spectrum, thereby extracting the pilot signal from the image transmission signal; Performing spectrum cropping on the pilot signal to obtain an enhanced pilot signal; The enhanced pilot signal is input into the pre-trained UAV model classification model to obtain the preliminary recognition result output by the UAV model classification model; The final drone model determination result is obtained by jointly judging multiple preliminary identification results through voting.

2. The lightweight and real-time OcuSync protocol drone identification method based on preamble extraction according to claim 1 is characterized in that: The collected drone signal is converted into a time-frequency spectrum through STFT, and the image transmission signal is extracted from the time-frequency spectrum using a double-window sliding threshold judgment algorithm based on the time-frequency spectrum, including: Initialize the first intermediate variable η left , the second intermediate variable η mid and the third intermediate variable η right , and preset the signal change trend judgment threshold η threld , the time length threshold T of the image transmission signal threld , and bandwidth threshold B threld ; Initialize the variable T at the start time of the recording signal start , variable T at the end time of the recording signal end , variable F that records the starting position of the signal on the spectrum axis start , and the variable F that records the end position of the signal on the spectrum axis end ; The collected drone signal of length FFTsize is converted into a time-frequency spectrum through STFT. The spectrum maximum of the time-frequency spectrum is extracted and sent to a sliding window of length L. If the sliding window is not full, the spectrum maximum is added to the empty space closest to the leftmost side of the sliding window. If the sliding window is full, the value on the leftmost side of the sliding window is removed and the new maximum value is added to the rightmost side of the sliding window. Whenever the sliding window is full, the energy value E of the left half of the sliding window is calculated left and the energy value E of the right half right , and calculate the energy ratio η, η=E right / E left , η mid Update to η left , η right Update to η mid , update η to η right ; According to η left ,η mid and η right The relationship between the size of η is used to judge the changing trend of η. If η right >η mid >η left , then it is determined that η continues to increase. If η right <η mid <η left , then it is determined that η continues to decrease; When T start =0, determine that η continues to increase and η>η threld When the signal starts, find the starting time point T start , start accumulating spectrum data, when it is determined that η continues to decrease and η<1 / η threld When the signal ends, find the end time point T end , stop accumulating spectrum data; Using the spectrum data sequence, repeat the sliding window operation to find the starting point F of the signal on the spectrum axis. start and end point F end , and according to T start 、T end 、F start and F end , calculate the time length T of the signal x and bandwidth B x ; If T x >T threld And B x >B threld , it is confirmed that the image transmission signal is detected, and according to T start and T end , extract the image transmission signal from the time-frequency spectrum.

3. The lightweight and real-time OcuSync protocol drone identification method based on preamble extraction according to claim 2 is characterized in that: The pilot signal and non-pilot signal in the image transmission signal are distinguished by using the margin factor feature based on the cross-correlation spectrum, thereby extracting the pilot signal from the image transmission signal, including: Consider each line of the image transmission signal as a time segment of the smallest unit signal, and calculate the normalized cross-correlation function between two adjacent time segments. Suppose the total number of time segments is M, and the time segment s is i With the previous time segment s i-1 The normalized cross-correlation function between i (τ), R i (τ) is calculated by the following formula: Where τ represents the time delay, μ i Indicates s i The mean value, μ i-1 Indicates s i-1 The mean of i Indicates s i The standard deviation, σ i-1 Indicates s i-1 The standard deviation of , E(·) means to obtain the expectation; The normalized cross-correlation function corresponding to each time segment is stored in a cross-correlation matrix R matrix middle; Calculate the margin factor of the normalized cross-correlation function corresponding to each time segment respectively, and set the time segment s i The corresponding margin factor is MF i , MF i Calculated by the following formula: Among them, max(·) means to find the maximum value, LE means R i the length of (τ); The margin factors corresponding to each time segment are combined into a feature vector MF; The clustering algorithm is used to classify MF, and all margin factors are divided into two categories to obtain the classification labels corresponding to each time segment. The two categories represent high correlation areas and low correlation areas, respectively. The clustering goal is to minimize the sum of squared errors within the class. The clustering goal is expressed by the formula: Among them, G represents the clustering target, c k represents the cluster center of the kth category, I(·) is the indicator function, if MF i Belongs to the kth category, then I(MF i ∈c k )=1, if MF i does not belong to the kth category, then I(MF i ∈c k )=0; Based on the margin factor, pulse detection is performed on each time segment to obtain the pulse detection result corresponding to each time segment. Let the time segment s i The corresponding pulse detection result is P i , P i Calculated by the following formula: Among them, max(c1,c2)-1 represents the preset clustering decision threshold, P i =1 means time segment s i The corresponding signal is the pilot signal, P i =0 means time segment s i The corresponding signal is a non-pilot signal; The pulse detection results corresponding to each time segment are combined into a pulse sequence P, where P i =1 corresponds to the position index corresponding to the time point of the pilot signal in the image transmission signal; Starting from the first pilot signal detected, the time-frequency spectra of N consecutive pilot signals are combined to obtain a pilot symbol; Each pilot symbol is filtered, and pilot symbols whose actual occupied number of rows is not equal to N are removed, and finally the pilot signal is extracted from the image transmission signal.

4. The lightweight and real-time OcuSync protocol drone identification method based on preamble extraction according to claim 3 is characterized in that: Perform spectrum cropping on the pilot signal to obtain an enhanced pilot signal, including: The pilot signal is converted to the frequency domain through DFT to obtain the spectrum of the pilot signal, and then the spectrum of the pilot signal is preprocessed including spectrum smoothing and spectrum binarization to obtain the preprocessed spectrum; Detect the widest pulse region in the pre-processed spectrum and determine the starting frequency point f of the widest pulse region start and the end frequency point f end , and calculate the center frequency point f center , f center =(f start +f end ) / 2; with f center As the center, the spectrum of the pilot signal is trimmed according to the preset retention width to obtain the trimmed spectrum R cropped (f), R cropped (f) is expressed by the formula: R cropped (f)=R(f)for f∈[f center -Δf,f center +Δf]; Wherein, R(f) represents the frequency spectrum of the pilot signal, and Δf represents the preset retention width; R cropped (f) Convert to the time domain through IDFT to obtain the enhanced pilot signal.

5. The lightweight and real-time OcuSync protocol drone identification method based on preamble extraction according to claim 1 is characterized in that: The network structure of the pre-trained drone model classification model is UAVResNet, which is improved on the basis of ResNet. It consists of a Layer0, a Layer1, an adaptive average pooling layer, and an output layer. Layer 0 consists of a 3×3 convolutional layer and a 3×3 maximum pooling layer. The 3×3 convolutional layer is followed by a BN normalization layer and a ReLU activation function layer. Layer 1 consists of a Bottleneck structure with 64 output channels. The Bottleneck structure consists of two 1×1 convolutional layers, a 3×3 convolutional layer, and a residual connection. The adaptive average pooling layer is used to resize the feature map output by Layer1 to 1×1; The output layer consists of a fully connected layer and a Softmax layer, which is used to perform classification based on the output of the adaptive average pooling layer and finally output the classification probability.

6. The lightweight and real-time OcuSync protocol drone identification method based on preamble extraction according to claim 5 is characterized in that: The pre-trained drone model classification model is trained through the following steps: Collect a large number of drone pilot signals, perform spectrum cropping on the collected drone pilot signals, and obtain enhanced drone pilot signals; The enhanced UAV pilot signal is used as a training sample and annotated with the true label to obtain a training sample set; Build UAVResNet as the initial model; The training samples in the training sample set are input into the initial model, and the initial model is iteratively trained until convergence to obtain a trained drone model classification model.

7. A lightweight and real-time OcuSync protocol drone identification method based on preamble extraction according to any one of claims 1 to 6, characterized in that: The final drone model determination result is obtained by voting on multiple preliminary identification results, including: Determining the confidence level of each preliminary recognition result based on the signal-to-noise ratio of the pilot signal corresponding to each preliminary recognition result, wherein the higher the signal-to-noise ratio of the pilot signal corresponding to the preliminary recognition result, the higher the confidence level of the preliminary recognition result; A weighted vote is performed based on the confidence of each preliminary recognition result to obtain the final drone model determination result.

8. A lightweight and real-time OcuSync protocol drone identification system based on preamble extraction, characterized by: The system comprises: Radio frequency receiving equipment, used to collect drone signals; The image transmission signal extraction module is used to convert the collected drone signal into a time-frequency spectrum through STFT, and use a dual-window sliding threshold judgment algorithm based on the time-frequency spectrum to extract the image transmission signal from the time-frequency spectrum; A pilot signal extraction module is used to distinguish pilot signals from non-pilot signals in the image transmission signal by using a margin factor feature based on a cross-correlation spectrum, thereby extracting the pilot signal from the image transmission signal; A data enhancement module is used to perform spectrum cropping on the pilot signal to obtain an enhanced pilot signal; A preliminary recognition module is used to input the enhanced pilot signal into a pre-trained UAV model classification model to obtain a preliminary recognition result output by the UAV model classification model; The joint judgment module is used to make a joint judgment on multiple preliminary recognition results through voting to obtain the final drone model judgment result.

9. An electronic device, characterized in that: include: at least one processor; and, a memory communicatively coupled to the at least one processor; In which, the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a lightweight and real-time OcuSync protocol drone identification method based on preamble code extraction as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it is possible to implement a lightweight and real-time OcuSync protocol drone identification method based on preamble code extraction as described in any one of claims 1 to 7.

Images