Unmanned aerial vehicle type and number identification method, system and device based on radio frequency signals and medium

Abstract

A radio frequency signal-based unmanned aerial vehicle type and number identification method, system and device, and a medium, the method comprising: preprocessing the signal of each unmanned aerial vehicle category in a compound database, and inputting the preprocessed signal into a radio frequency feature identification model of multiple unmanned aerial vehicle categories for training; preprocessing a to-be-detected unmanned aerial vehicle signal, and inputting the preprocessed to-be-detected unmanned aerial vehicle signal into the trained radio frequency feature recognition model of the unmanned aerial vehicle category to obtain a to-be-detected unmanned aerial vehicle category label; performing cross-correlation operation on the to-be-detected unmanned aerial vehicle signal and the standard reference signal corresponding to the to-be-detected unmanned aerial vehicle category label, and then performing peak detection to obtain the number of each type of unmanned aerial vehicles; according to the method, on the basis of high-precision recognition of a radio frequency feature recognition model of multiple unmanned aerial vehicle types, the cooperative effect of cross-correlation operation and peak detection is combined, the recognition accuracy of the multiple unmanned aerial vehicles is improved, and meanwhile accurate statistics of the multiple unmanned aerial vehicle types is achieved; the system, the equipment and the medium are used for implementing the method.

Description

Technical Field

[0001] This invention relates to the field of drone identification technology, specifically to a method, system, device, and medium for identifying the type and number of drones based on radio frequency signals. Background Technology

[0002] Currently, common drone detection technologies mainly include optical recognition, radar detection, and acoustic detection. Optical recognition is greatly affected by the environment, performs poorly at night, and is prone to missed detection due to obstruction; radar detection is an active detection method, which is costly; acoustic detection has a short range.

[0003] The technology of identifying drones using their radio frequency (RF) signatures is commonly referred to as radio frequency (RF) fingerprinting. When a drone interacts with the ground during flight, it emits specific RF signals. By capturing and analyzing these signals, the drone's model, manufacturer, or other specific characteristics can be identified. The advantages of RF fingerprinting include its electromagnetic properties being difficult to conceal, high reliability, no environmental impact, and strong passive detection stealth. However, it is susceptible to interference from other drones operating on the same frequency, and the equipment cost is relatively high.

[0004] Traditional radio frequency (RF) fingerprinting primarily involves signal acquisition, spectrum analysis, feature extraction, RF fingerprint database establishment, identification, and classification. First, specialized radio frequency (RF) signals from the surrounding environment are captured using specialized radio receiving equipment. These signals undergo spectrum analysis, breaking them down into different frequency components to help identify the frequency characteristics of different signals. Next, specific features are extracted from the analyzed signals. These features can include frequency, pulse mode, modulation method, etc. Since different drone models typically produce different RF characteristics, the source of the RF signal can be identified by comparing the extracted RF features with a pre-established RF fingerprint database. Finally, by combining the RF fingerprint information of known drone models and manufacturers stored in the database, the drone model, manufacturer, or other specific characteristics can be identified through comparison.

[0005] With the development of communication technology, the characteristics of frequency hopping sequences are becoming increasingly complex. Faced with increasingly complex frequency hopping communication, spread spectrum technology and dynamically changing signal environments, traditional identification methods based on manually designed features face many difficulties. For example, the feature representation capability is limited, and manually designed features are difficult to effectively capture high-dimensional, nonlinear, and time-varying complex signal patterns, especially signals generated by advanced technologies such as deep frequency hopping and adaptive modulation.

[0006] Deep learning-driven intelligent detection methods demonstrate significant advantages in complex signal environments. Deep neural networks can directly learn deep features from the original signal (IQ data) or its time-frequency representation (such as spectrograms and time spectra), and excel at capturing complex nonlinear relationships in signals. They can effectively cope with the effects of complex modulation, frequency hopping sequences, and dynamic environments. By training on diverse data, the model can learn more fundamental and robust feature representations, exhibiting better generalization ability to unknown devices and environmental changes. Therefore, deep learning-based intelligent radio frequency identification has become a new development direction in the field of drone detection, identification, and tracking.

[0007] Liu Bing (Liu Bing, Shi Mingxin, Liu Jiaqi. Research on UAV RF Signal Recognition Algorithm Based on Wavelet Entropy Features [J]. Journal of Electronics and Information Technology, 2025, 47(08):2736-2745.) proposed a UAV RF signal recognition method based on wavelet entropy features. This method collects RF signals of 6 common UAVs from the Drone RFA dataset, extracts wavelet entropy features to construct vectors, and optimizes the SVM classifier with the Big Cane Mouse Optimization Algorithm. This solves the problems of insufficient accuracy, low signal-to-noise ratio, and poor stability of traditional SVM in signal classification, resulting in a 10-fold cross-validation accuracy of over 97%. However, the drawback is that it can only recognize single UAV signals and cannot handle the situation of multiple signal aliasing and target interference.

[0008] The invention patent with publication number CN114358056A proposes a method and system for drone signal recognition based on deep learning. This method involves acquiring signals from known drone models using a receiver, generating a time-frequency map through short-time Fourier transform, constructing a sample library, and then using a YOLOv4 neural network to identify the drone model based on energy bars in the time-frequency map and visual differences. This method can accurately identify single drone signals under conditions of impulse interference, low signal-to-noise ratio, and some time-frequency aliasing. However, it lacks the ability to simultaneously identify multiple drone signals and does not fully consider the issues of mutual interference between multiple targets and the accuracy of feature extraction. Summary of the Invention

[0009] To overcome the shortcomings of the existing technology, the present invention aims to provide a method, system, device, and medium for identifying the type and number of drones based on radio frequency signals. This method, based on the high-precision feature recognition of radio frequency feature recognition models for multiple drone categories, combines the synergistic mechanism of cross-correlation operation with standard signals corresponding to category labels and peak detection. This solves the problem in the existing technology of "weak category differentiation ability and inability to accurately count the specific number of drones of each type in multi-drone overlapping scenarios". Thus, while improving the accuracy of multi-drone recognition, it also achieves high-precision and high-efficiency classification of the types of multiple drones, which can better adapt to the recognition needs of multi-drone overlapping in real-world scenarios.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for identifying the type and number of drones based on radio frequency signals, specifically including the following steps: Step 1: Preprocess the signals of each UAV category in the composite database to obtain a time-frequency image dataset of radio frequency signals of multi-target UAVs; Step 2: Input the time-frequency image dataset of radio frequency signals of multi-target UAVs into the radio frequency feature recognition model for multi-UAV categories for training; Step 3: The drone signal to be detected is preprocessed as described in Step 1. The time-frequency image of the radio frequency signal of the drone to be detected obtained after preprocessing is input into the trained radio frequency feature recognition model of the drone category to obtain the drone category label. Step 4: Perform cross-correlation calculation on the UAV signal to be detected and the standard reference signal corresponding to the UAV category label to be detected to obtain the UAV cross-correlation sequence; Step 5: Perform peak detection on the cross-correlation sequence of the drones to obtain the number of drones of each type.

[0011] Furthermore, step 1 specifically includes: Step 1.1: Sample the signals of each drone category in the composite database with a preset probability, and multiply the sampled signals of each drone category by a random coefficient to obtain the basic dataset and the extended dataset; Step 1.2: Normalize the aliased UAV radio frequency signals in the basic dataset and extended dataset using a linear transformation; Step 1.3: Truncate the normalized aliased UAV radio frequency signal to obtain the aliased UAV radio frequency truncated signal; Step 1.4: Perform a short-time Fourier transform on the aliased UAV radio frequency truncation signal to obtain the time-spectrum matrix; Step 1.5: Use a plotting tool to present the time-frequency matrix as a time-frequency image to obtain a time-frequency image dataset of radio frequency signals of multi-target UAVs.

[0012] Furthermore, the radio frequency feature recognition model for multiple UAV categories in step 2 includes a feature extraction network and a detection head; the feature extraction network adopts the Efficient Net architecture to map the input radio frequency signal time-frequency image into a discriminative feature vector space; the detection head consists of a fully connected layer to classify the extracted features to achieve feature-to-category mapping.

[0013] Furthermore, the feature extraction network is composed of a first convolutional layer, a normalization layer, a first activation layer, a moving inverse residual group, a second convolutional layer, a first batch of normalization layers, a second activation layer, and a pooling layer connected in series. The moving inverse residual group is composed of seven moving inverse residual layers with different repetition counts connected in series, and each moving inverse residual layer is connected to the others through a second batch of normalization layers and a first Swish activation layer. The moving inverse residual layer is composed of a third convolutional layer, a third batch of normalization layers, a third activation layer, a fourth convolutional layer, a fourth batch of normalization layers, a fourth activation layer, an SE attention layer, a fifth convolutional layer, a fifth batch of normalization layers, a Dropout layer, and a residual connection layer connected in series. The SE attention layer is composed of an average pooling layer, a second Swish activation layer, a Sigmoid activation layer, and a 1×1 convolutional layer connected in series.

[0014] Further, step 2 specifically includes: using a data loader to load the time-frequency image dataset of radio frequency signals of multiple target UAVs into the radio frequency feature recognition model of multiple UAV categories, using the binary cross-entropy loss function as the loss function of the radio frequency feature recognition model of multiple UAV categories, and then using an adaptive moment estimation optimizer to iteratively update the weight values ​​of the network to achieve the convergence of the radio frequency feature recognition model of multiple UAV categories.

[0015] Furthermore, in step 4, the standard reference signal corresponding to the drone category label to be detected is obtained by matching the drone category label to be detected with the category labels in the composite database.

[0016] Furthermore, step 5 specifically includes: During peak detection, record the maximum peak value and the position of the maximum peak value. And set a detection threshold for the number of drones, then start from the position of the maximum peak. Peak detection is performed on both sides of the peak. If there is a secondary peak that exceeds the threshold and the time delay difference between the secondary peak that exceeds the threshold and the maximum peak is greater than the set sampling point, it is determined to be an independent drone signal. The drone type of the independent drone signal corresponds to the drone type of the standard reference signal. After the drone type of the independent drone signal is confirmed, the drone count counter corresponding to that type is incremented by 1 until the sampling points on both sides of the peak are scanned to obtain the number of drones of each type.

[0017] A drone type and quantity identification system based on radio frequency signals, comprising: Data preprocessing module: preprocesses the signals of each UAV category in the composite database to obtain a time-frequency image dataset of radio frequency signals of multi-target UAVs; Model training module: Input the time-frequency image dataset of radio frequency signals of multi-target UAVs into the radio frequency feature recognition model of multi-UAV categories for training; Drone Category Label Recognition Module: The data preprocessing module preprocesses the drone signal to be detected, and inputs the preprocessed time-frequency image of the drone radio frequency signal to be detected into the trained radio frequency feature recognition model of drone category to obtain the drone category label to be detected. Cross-correlation module: Performs cross-correlation calculations on the UAV signal to be detected and the standard reference signal corresponding to the UAV category label to be detected, and obtains the cross-correlation sequence of the UAV; Peak detection module: Performs peak detection on the cross-correlation sequence of UAVs to obtain the number of UAVs of each type.

[0018] A device for identifying the type and number of drones based on radio frequency signals, comprising: Memory: Used to store computer programs that implement the radio frequency signal-based method for identifying the number of UAV types as described above; Processor: Used to implement the method for identifying the type and number of drones based on radio frequency signals as described above when executing the computer program.

[0019] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for identifying the type and number of unmanned aerial vehicles based on radio frequency signals as described above.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a multi-drone category RF feature recognition model based on the EfficientNet architecture, solving the problems of insufficient feature extraction and poor category discrimination ability of existing single-classification models for multi-target aliased RF signals. The EfficientNet architecture, through a composite scaling strategy of "depth, width, and resolution," can capture subtle features of different drone signals in time-frequency images at a lower computational cost. Compared with traditional models that can only extract features in a single dimension, the feature extraction dimensions are more comprehensive. At the same time, the multi-drone category RF feature recognition model of this invention supports parallel recognition of multiple drone categories in aliased time-frequency images, rather than the single-category output of traditional models. In simulation tests, the macro average accuracy reaches 96.56% on the basic dataset and 97.65% on the extended dataset, showing high feature extraction efficiency, strong compatibility for multi-category recognition, and good scalability.

[0021] 2. This invention, based on the prediction of drone category labels using a multi-drone category RFID recognition model, uses cross-correlation calculations and peak detection to statistically determine the specific number of each type of drone, solving the problem of existing technologies that "can only count the specific number of drones, but cannot count drone types." Theoretically, if there are N drones of the same type, the cross-correlation result will show N obvious peaks. Combining peak detection with this result can accurately obtain the specific number of each type of drone. This process does not require additional ranging equipment, reducing the deployment cost and complexity of the entire system and improving its stability and reliability. Furthermore, based on the distribution of the confusion matrix in simulation experiments, the drone identification accuracy of this invention reaches 96.1%, and the drone classification accuracy reaches 94.3%, demonstrating high detection accuracy and high efficiency.

[0022] In summary, this invention, based on the high-precision feature recognition of radio frequency feature recognition models for multiple drone categories, combines a synergistic mechanism of cross-correlation calculation of standard signals corresponding to category labels and peak detection to solve the problems of "weak category differentiation ability and inability to accurately count the specific number of each type of drone in multi-drone overlapping scenarios" in the prior art. Thus, while improving the accuracy of multi-drone recognition, it also achieves high-precision and high-efficiency classification of the types of multiple drones, and can better adapt to the recognition needs of multi-drone overlapping scenarios in real-world scenarios. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method for identifying the type and number of drones based on radio frequency signals according to the present invention.

[0024] Figure 2 This is a schematic diagram of the structure of the radio frequency feature recognition model for multiple drone categories of the present invention.

[0025] Figure 3 This is a comparison chart showing the accuracy of the basic architecture of the multi-drone category radio frequency feature recognition model of the present invention with that of existing basic architectures.

[0026] Figure 4 This is the confusion matrix for each of the six drone categories in the basic dataset of this invention.

[0027] Figure 5 This is the confusion matrix for each of the twelve drone categories in the extended dataset of this invention.

[0028] Figure 6 This is the confusion matrix for multi-UAV number classification and recognition in this invention. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments: See Figure 1 A method for identifying the type and number of drones based on radio frequency signals, specifically including the following steps: Step 1: Preprocess the signals of each UAV category in the composite database to obtain a time-frequency image dataset of radio frequency signals of multi-target UAVs; the composite database mainly includes the Drone RFA dataset and the Drone RF dataset.

[0030] The Drone RFA dataset is a large dataset containing 24 types of drones and 1 type of background signal. The sampling frequency is 100MHz. It includes both commercial and civilian drones. Each type of drone has at least 12 sampling segments, and each segment has no less than 100 million sampling points. For outdoor signals, there are sampling segments at three different distances: far, medium, and near. The Drone RF dataset contains 227 radio frequency data segments collected in complex electromagnetic environments from three types of drones. The dataset includes recorded segments of drone radio frequency background activity, as well as operational segments of drones in different modes (such as on, off, hovering, flying, and video recording). The dataset contains more than 40GB of drone radio frequency signal data, covering various radio frequency characteristics. Since the Drone RFA dataset contains a large number of drone radio frequency segments, has a complete range of types, and covers a wide range, and was published in 2024, this invention mainly uses the Drone RFA dataset as the primary data source, while using the Drone RF dataset as an auxiliary verification.

[0031] In order to simulate the situation where multiple drones exist simultaneously in space, this embodiment can generate simulation data of radio frequency signals of multiple drones through preprocessing, thereby simulating the time-frequency overlap phenomenon that may occur when multiple drones work at the same time in a real scenario.

[0032] Step 1 specifically includes: Step 1.1: Sample the signals of each drone category in the composite database with a preset probability, and multiply the sampled signals of each drone category by a random coefficient to obtain the basic dataset and the extended dataset; in this embodiment, the preset sampling probability is 33.3%, that is, the probability of each drone category being sampled is 33.3%, and the random coefficient is 0.1~1.0; The basic dataset contains 6 types of drone signals and background signals. Each type of signal has a sampling probability of 33.3%, and the expected number of drone categories is 2. The actual sampling results may include all 6 drone categories, or they may only include background signals without drone signals. The extended dataset contains 12 types of drone signals, with each type of signal having a sampling probability of 33.3%, and the expected number of drone types is 4. In terms of signal strength processing, the first sampled signal retains its original amplitude, while subsequent signals are multiplied by a random coefficient in the range of 0.1 to 1.0 to simulate the signal strength differences in real-world scenarios.

[0033] Step 1.2: Apply a linear transformation to normalize the aliased UAV radio frequency signals in the basic dataset and the extended dataset, so that the amplitude of the aliased UAV radio frequency signals is linearly mapped to the [0,1] interval; The specific expression for the linear transformation is as follows: in, It is the normalized value. These are signal points that need to be normalized. This represents the minimum point in the signal. This indicates the maximum value point in the signal; Step 1.3: Truncate the normalized aliased drone radio frequency signal to obtain the aliased drone radio frequency truncated signal; In this embodiment, the normalized aliased drone radio frequency signal is truncated to 0.1 seconds. Since the original dataset sampling rate is 100M, the length of the obtained aliased drone radio frequency truncated signal is 10M. Step 1.4: Perform a short-time Fourier transform on the aliased UAV radio frequency truncation signal to obtain the time-frequency spectrum matrix; the specific expression of the short-time Fourier transform is as follows: in, The signal to be analyzed For window functions, This is the time-frequency spectrum matrix after the short-time Fourier transform; In this embodiment, the step size of the short-time Fourier transform is set to 1024, and the window function is set to a Hamming window. The Hamming window function Specifically: in, For window length, This represents the i-th sampling point in the sequence; the window length in this embodiment is... Set it to 2048.

[0034] Step 1.5: Use a plotting tool to present the time-spectrum matrix as a time-frequency image, obtaining a time-frequency image dataset of the radio frequency signal of the multi-target UAV. In this embodiment, Python is used as the plotting tool; functions in Python's matplotlib library can present the time-spectrum matrix as a time-frequency image.

[0035] Step 2: Input the time-frequency image dataset of radio frequency signals of multi-target UAVs into the radio frequency feature recognition model for multi-UAV categories for training; See Figure 2 The multi-drone category radio frequency feature recognition model in step 2 includes a feature extraction network and a detection head. The feature extraction network adopts the Efficient Net architecture to map the input radio frequency signal time-frequency image to a discriminative feature vector space. The detection head consists of a fully connected layer to classify the extracted features to achieve feature-to-category mapping. The feature extraction network converts the radio frequency signal time-frequency image into a feature vector, which is then input into the fully connected layer of the detection head. The output is a vector with a length equal to the number of drone categories. The output of each category is then compressed to the (0,1) interval by the Sigmoid activation function as an estimate of the probability of existence of each category.

[0036] The feature extraction (Efficient Net) network consists of a first convolutional layer, a normalization layer, a first activation layer, a moving inverse residual group, a second convolutional layer, a first batch of normalization layers, a second activation layer, and a pooling layer connected in series. Referring to Table 1, the moving inverse residual group consists of seven moving inverse residual layers (MBConv) with different repetition counts connected in series, with each moving inverse residual layer connected via a second batch of normalization layers and a first Swish activation layer. The moving inverse residual layer consists of a third convolutional layer, a third batch of normalization layers, a third activation layer, a fourth convolutional layer, a fourth batch of normalization layers, a fourth activation layer, an SE attention layer, a fifth convolutional layer, a fifth batch of normalization layers, a Dropout layer, and a residual connection layer connected in series. The SE attention layer consists of an average pooling layer, a second Swish activation layer, a Sigmoid activation layer, and a 1×1 convolutional layer connected in series.

[0037] In this embodiment, the most important part of the EfficientNet network is the moving inverse residual group. Each moving inverse residual layer in the moving inverse residual group reduces the amount of computation and parameters through depthwise separable convolution and inverted residual structure, making the model occupy less memory during training and inference, which helps the EfficientNet model to be more easily deployed on resource-constrained devices. The SE attention layer can enhance the feature channels that are useful for information and suppress the feature channels that are not important to the current task, thereby improving the representation ability of the network.

[0038] Table 1 below shows the main structural configuration of the feature extraction network: Table 1. Main structural configuration of feature extraction network Step 2 specifically includes: using a data loader to load the time-frequency image dataset of radio frequency signals from multiple target drones into a multi-drone category radio frequency feature recognition model; using the binary cross-entropy loss function as the loss function of the multi-drone category radio frequency feature recognition model; and then iteratively updating the network weights using an adaptive moment estimation (Adam) optimizer to increase the distance between images of different categories in the feature space, thereby achieving convergence of the multi-drone category radio frequency feature recognition model. Simultaneously, the experimental tracking platform (Wandb) is used to monitor the loss function and accuracy during training.

[0039] In multi-label classification tasks, each sample may correspond to multiple drone categories simultaneously, making it difficult to construct effective sample combinations. Single-label tasks use center loss, which requires each class to have a unique and stable center point. However, due to the more complex feature space distribution of multi-label samples, it is impossible to reasonably select class centers. Therefore, in multi-label classification tasks, a more suitable loss function is needed. This embodiment uses binary cross-entropy loss for multi-class classification, which is independent for each class prediction. The expression for the binary cross-entropy loss function is: in, For the actual label vector, This is the model's predicted output vector. This represents the total number of categories for the classification task.

[0040] The above labels do not use one-hot encoding, but rather multi-hot encoding. If a sample belongs to the first... Class and the Class, then the first in the label vector The position and the first The label is 1 for the first position and 0 for the other positions. The label for the background sample (which does not contain any drone signals) is also 0.

[0041] During the inference phase, the radio frequency feature recognition model for multiple drone categories in this embodiment will output a prediction vector with a length equal to the number of all drone categories. Unlike single-label classification, if the predicted probability of each category is greater than a preset threshold (i.e., a confidence threshold), then the drone signal of that category is considered to exist, and the corresponding label is set to 1; otherwise, it is set to 0, thereby obtaining the category label prediction results for multiple drone categories. This method gives the multi-label classification task greater flexibility and helps to improve the recognition ability when multiple drone signals appear simultaneously.

[0042] Step 3: The drone signal to be detected is preprocessed as described in Step 1. The time-frequency image of the radio frequency signal of the drone to be detected obtained after preprocessing is input into the trained radio frequency feature recognition model of the drone category to obtain the drone category label. When the category label of the drone to be detected is known, the number of drones of the same type can be accurately detected through cross-correlation operation and peak detection. Since the radio frequency signals of drones of the same category have similar modulation characteristics and frequency hopping sequences, the time delay difference between signals can be characterized by the separation degree of cross-correlation peaks.

[0043] Step 4: Perform cross-correlation calculation on the UAV signal to be detected and the standard reference signal corresponding to the UAV category label to be detected to obtain the UAV cross-correlation sequence; In step 4, the standard reference signal corresponding to the drone category label to be detected is obtained by matching the drone category label to be detected with the category labels in the composite database; the expression for the cross-correlation operation is: Where, in the formula In order to receive signals, For reference signal, For time delay variables; Step 5: Perform peak detection on the cross-correlation sequence of the drones to obtain the number of drones of each type.

[0044] Step 5 specifically includes: During peak detection, record the maximum peak value and the position of the maximum peak value. And set a detection threshold for the number of drones, then start from the position of the maximum peak. Peak detection is performed on both sides of the peak. If there is a secondary peak that exceeds the threshold and the time delay difference between the secondary peak that exceeds the threshold and the maximum peak is greater than the set sampling point, it is determined to be an independent drone signal. The drone type of the independent drone signal corresponds to the drone type of the standard reference signal. After the drone type of the independent drone signal is confirmed, the drone count counter corresponding to that type is incremented by 1 until the sampling points on both sides of the peak are scanned to obtain the number of drones of each type.

[0045] The drone number detection threshold set in this embodiment is: (i.e., 0.8 times the maximum peak value) to avoid false peaks, the number of sampling points is set to 1000.

[0046] A drone type and quantity identification system based on radio frequency signals, comprising: Data preprocessing module: preprocesses the signals of each UAV category in the composite database to obtain a time-frequency image dataset of radio frequency signals of multi-target UAVs; Model training module: Input the time-frequency image dataset of radio frequency signals of multi-target UAVs into the radio frequency feature recognition model of multi-UAV categories for training; Drone Category Label Recognition Module: The data preprocessing module preprocesses the drone signal to be detected, and inputs the preprocessed time-frequency image of the drone radio frequency signal to be detected into the trained radio frequency feature recognition model of drone category to obtain the drone category label to be detected. Cross-correlation module: Performs cross-correlation calculations on the UAV signal to be detected and the standard reference signal corresponding to the UAV category label to be detected, and obtains the cross-correlation sequence of the UAV; Peak detection module: Performs peak detection on the cross-correlation sequence of UAVs to obtain the number of UAVs of each type.

[0047] A device for identifying the type and number of drones based on radio frequency signals, comprising: Memory: Used to store computer programs that implement the radio frequency signal-based method for identifying the number of UAV types as described above; Processor: Used to implement the method for identifying the type and number of drones based on radio frequency signals as described above when executing the computer program.

[0048] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for identifying the type and number of unmanned aerial vehicles based on radio frequency signals as described above.

[0049] The application effects of this invention will be described in detail below with reference to simulation experiments.

[0050] Simulation test environment configuration: Windows 11 operating system is used, and a development framework including Python 3.7.1 interpreter is built. CUDA 11.6 parallel computing architecture and cuDNN acceleration library are selected to achieve GPU hardware acceleration. The deep learning framework is based on PyTorch 1.13.0 and its matching Torchvision 0.14.0 computer vision library. This combination of versions ensures the flexibility of dynamic construction of the computation graph while giving full play to the mixed-precision computing capabilities of NVIDIA RTX 30 / 40 series graphics cards.

[0051] 1. The simulation first tests the ability of the multi-UAV category RF feature recognition model of this invention to perform binary classification detection of the presence of UAV RF signals. The core is to distinguish between target signals (UAV RF features) and environmental interference signals; the experimental design is as follows: The positive sample group contains time-frequency images of radio frequency signals from eight types of commercial drones, including DJI Phantom 3 and DJI Mavic Pro, some of which contain interference signals, totaling 2000 time-frequency images. The negative sample group contains only interference signals such as Wi-Fi and Bluetooth, totaling 500 time-frequency images. The training and test sets are divided in a 3:2 ratio. EfficientNet-b3 is used as the backbone network for transfer learning, and the Adam optimizer is used. The number of training epochs is set to 30. After the model converges, the drone signal recognition and detection results are shown in Table 2. Table 2. Identification and Detection of UAV Signals As shown in Table 2, the recognition accuracy of the radio frequency feature recognition model for multiple UAV categories of the present invention is 99.6%, which is excellent in binary classification tasks. The UAV signal can be accurately detected by the time-frequency diagram of the UAV radio frequency signal, and the false alarm rate is extremely low.

[0052] 2. The basic architecture of the multi-UAV category radio frequency feature recognition model of this invention is the Efficient Net series, which includes Efficient Net-b0, Efficient Net-b3, and Efficient Net-b5. The effectiveness of the basic architecture model is verified by comparing it with existing VGG16, AlexNet, and ResNet-50 models. The experimental design is as follows: The experimental dataset contains 2000 time-frequency images of radio frequency signals from eight types of commercial drones. These images are divided into training and testing sets in a 3:2 ratio. All pre-trained basic architecture models were fine-tuned using the PyTorch framework, with a uniform training cycle of 60 epochs. The Adam optimizer was used, and a training strategy of decreasing the learning rate by 50% every 20 epochs was implemented. The results are as follows: Figure 3The test accuracy of the different infrastructure models shown; from Figure 3 It can be concluded that the EfficientNet series exhibits significant advantages, with the EfficientNet-b0, EfficientNet-b3, and EfficientNet-b5 versions achieving accuracies of 96.9%, 97.3%, and 97.2%, respectively. All of these outperform the traditional VGG16, AlexNet, and ResNet-50 network architectures. In particular, EfficientNet-b3 achieves the best balance between model complexity and classification performance, with its 97.3% accuracy improving upon the second-best ResNet-50 architecture by 3.1 percentage points.

[0053] 3. In the simulation of the radio frequency feature recognition model for multiple UAV categories, the training and test sets of the basic dataset were sampled 600 times and 400 times respectively. The UAV signals in the two sampled datasets do not overlap at all, thus ensuring the independence and rigor of the experimental data. For the extended dataset, due to the high complexity of the signal images, the training convergence speed is slow. Therefore, the training and test sets of the extended dataset were sampled 3200 times and 800 times respectively, and the number of training rounds was increased to ensure that the model can fully learn the features of each category.

[0054] Due to the inherent characteristics of multi-label classification, a misclassification of a label does not necessarily mean that the label was misclassified into another category. Therefore, traditional confusion matrices are insufficient to intuitively represent the distribution characteristics of classification errors. To more clearly analyze the model's classification performance across different categories, binary confusion matrices were plotted for each category. Figure 4 The classification results of each category in the basic dataset are shown. After the radio frequency feature recognition model of the multi-drone category of this invention is extended from binary classification task to multi-label task, the overall classification ability does not show significant degradation. As the number and types of drones in the sample increase, the classification accuracy of each category does not decrease significantly. However, the correct classification rate of the overall sample is affected, which is reflected in the decrease in accuracy.

[0055] Referring to Table 3, during the training and optimization process of the extended dataset, the Hamming loss was significantly reduced and the macro average accuracy was improved. This indicates that the metric at the label level did not decrease with the increase in the types and number of drones, but rather improved, demonstrating that the radio frequency feature recognition model for multiple drone categories in this invention has strong scalability.

[0056] Table 3. Radio Frequency Identification Indicators for Multiple UAV Categories in the Basic and Extended Datasets like Figure 5As shown, to intuitively analyze the classification performance of the multi-drone category RFID model of this invention across different categories, a confusion matrix of 12 categories in the extended dataset was plotted. When the number of categories is expanded to 12, the recognition ability of the multi-drone category RFID model of this invention for each category does not show a significant decrease. As the number of categories increases, the correct classification of complete samples becomes more stringent, resulting in a decrease in sample-level accuracy to 86.50%. However, it is worth noting that the extended dataset provides more samples and increases the number of training rounds. Therefore, from the perspective of label-level evaluation metrics, the performance of the multi-drone category RFID model of this invention is still improved. This indicates that the multi-drone category RFID model of this invention, limited by data scale and computing resources, has not reached the optimal performance of the model, but it exhibits strong scalability.

[0057] 4. In the simulation of multi-drone number recognition, since it is difficult to obtain different numbers of the same type of drones in reality, a construction method similar to that used in the multi-category drone simulation above was adopted to simulate the signals of multiple drones of the same type in space. Signals of 0 to 5 drones of the same type were superimposed. To simulate different spatial distances, a certain attenuation coefficient was added to each signal, and a multi-drone number recognition dataset was constructed. The samples in this dataset were generated as follows: First, a single type of drone signal was selected and sampled 0 to 5 times, then superimposed onto the background signal. Each signal superposition involved a random time delay to simulate the difference in the frequency hopping sequence. The total number of samples in the dataset was 300, and the total number of drone signals designed was 750 (i.e., 50 samples from each of 0 to 5 drones). Each group lasted 10ms, and the sampling rate was 100MHz.

[0058] like Figure 6 As shown, by detecting the multi-drone number recognition dataset, 721 out of 750 drone signals were detected, and 29 drone signals were not detected. The number of drones identified accounted for 96.1% of all drones. If the number of drones is regarded as the classification label, there are 17 incorrectly identified samples. The total number of samples is 300, and the classification accuracy rate is 94.3%.

[0059] The simulation results show that, using the method of the present invention, the specific number of each type of drone can be accurately detected when the target drone category is known.

[0060] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. For those skilled in the art, any modifications, equivalent substitutions or improvements made within the spirit of the technical solution of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying the type and number of drones based on radio frequency signals, characterized in that: Specifically, the following steps are included: Step 1: Preprocess the signals of each UAV category in the composite database to obtain a time-frequency image dataset of radio frequency signals of multi-target UAVs; Step 2: Input the time-frequency image dataset of radio frequency signals of multi-target UAVs into the radio frequency feature recognition model for multi-UAV categories for training; Step 3: The drone signal to be detected is preprocessed as described in Step 1. The time-frequency image of the radio frequency signal of the drone to be detected obtained after preprocessing is input into the trained radio frequency feature recognition model of the drone category to obtain the drone category label. Step 4: Perform cross-correlation calculation on the UAV signal to be detected and the standard reference signal corresponding to the UAV category label to be detected to obtain the UAV cross-correlation sequence; Step 5: Perform peak detection on the cross-correlation sequence of the drones to obtain the number of drones of each type.

2. The method for identifying the type and number of drones based on radio frequency signals according to claim 1, characterized in that: Step 1 specifically includes: Step 1.1: Sample the signals of each drone category in the composite database with a preset probability, and multiply the sampled signals of each drone category by a random coefficient to obtain the basic dataset and the extended dataset; Step 1.2: Normalize the aliased UAV radio frequency signals in the basic dataset and extended dataset using a linear transformation; Step 1.3: Truncate the normalized aliased UAV radio frequency signal to obtain the aliased UAV radio frequency truncated signal; Step 1.4: Perform a short-time Fourier transform on the aliased UAV radio frequency truncation signal to obtain the time-spectrum matrix; Step 1.5: Use a plotting tool to present the time-frequency matrix as a time-frequency image to obtain a time-frequency image dataset of radio frequency signals of multi-target UAVs.

3. The method for identifying the type and number of drones based on radio frequency signals according to claim 1, characterized in that: The radio frequency feature recognition model for multiple UAV categories in step 2 includes a feature extraction network and a detection head; the feature extraction network adopts the Efficient Net architecture, which is used to map the time-frequency image of the input radio frequency signal to a discriminative feature vector space; The detection head consists of a fully connected layer, which is used to classify the extracted features to achieve feature-to-category mapping.

4. The method for identifying the type and number of drones based on radio frequency signals according to claim 3, characterized in that: The feature extraction network is composed of a first convolutional layer, a normalization layer, a first activation layer, a moving inverse residual group, a second convolutional layer, a first batch of normalization layers, a second activation layer, and a pooling layer connected in series. The moving inverse residual group is composed of seven moving inverse residual layers with different repetition counts connected in series, and each moving inverse residual layer is connected to the other through a second batch of normalization layers and a first Swish activation layer. The moving inverse residual layer is composed of a third convolutional layer, a third batch of normalization layers, a third activation layer, a fourth convolutional layer, a fourth batch of normalization layers, a fourth activation layer, an SE attention layer, a fifth convolutional layer, a fifth batch of normalization layers, a Dropout layer, and a residual connection layer connected in series. The SE attention layer is composed of an average pooling layer, a second Swish activation layer, a Sigmoid activation layer, and a 1×1 convolutional layer connected in series.

5. A method for identifying the type and number of drones based on radio frequency signals according to claim 1 or 3, characterized in that: Step 2 specifically includes: using a data loader to load the time-frequency image dataset of radio frequency signals of multiple target UAVs into the radio frequency feature recognition model of multiple UAV categories, using the binary cross-entropy loss function as the loss function of the radio frequency feature recognition model of multiple UAV categories, and then using an adaptive moment estimation optimizer to iteratively update the weight values ​​of the network to achieve the convergence of the radio frequency feature recognition model of multiple UAV categories.

6. The method for identifying the type and number of drones based on radio frequency signals according to claim 1, characterized in that: In step 4, the standard reference signal corresponding to the drone category label to be detected is obtained by matching the drone category label to be detected with the category labels in the composite database.

7. The method for identifying the type and number of drones based on radio frequency signals according to claim 1, characterized in that: Step 5 specifically includes: During peak detection, record the maximum peak value and the position of the maximum peak value. And set a detection threshold for the number of drones, then start from the position of the maximum peak. Peak detection is performed on both sides of the peak. If there is a secondary peak that exceeds the threshold and the time delay difference between the secondary peak that exceeds the threshold and the maximum peak is greater than the set sampling point, it is determined to be an independent drone signal. The drone type of the independent drone signal corresponds to the drone type of the standard reference signal. After the drone type of the independent drone signal is confirmed, the drone count counter corresponding to that type is incremented by 1 until the sampling points on both sides of the peak are scanned to obtain the number of drones of each type.

8. A drone type and quantity identification system based on radio frequency signals, characterized in that: include: Data preprocessing module: preprocesses the signals of each UAV category in the composite database to obtain a time-frequency image dataset of radio frequency signals of multi-target UAVs; Model training module: Input the time-frequency image dataset of radio frequency signals of multi-target UAVs into the radio frequency feature recognition model of multi-UAV categories for training; Drone Category Label Recognition Module: The data preprocessing module preprocesses the drone signal to be detected, and inputs the preprocessed time-frequency image of the drone radio frequency signal to be detected into the trained radio frequency feature recognition model of drone category to obtain the drone category label to be detected. Cross-correlation module: Performs cross-correlation calculations on the UAV signal to be detected and the standard reference signal corresponding to the UAV category label to be detected, and obtains the cross-correlation sequence of the UAV; Peak detection module: Performs peak detection on the cross-correlation sequence of UAVs to obtain the number of UAVs of each type.

9. A device for identifying the type and number of drones based on radio frequency signals, characterized in that: include: Memory: Used to store computer programs implementing the method for identifying the type and number of unmanned aerial vehicles based on radio frequency signals as described in any one of claims 1-7; Processor: Used to implement the method for identifying the type and number of unmanned aerial vehicles based on radio frequency signals as described in any one of claims 1-7 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method for identifying the type and number of unmanned aerial vehicles based on radio frequency signals as described in any one of claims 1-7.

Images