Unmanned aerial vehicle radio frequency signal detection and identification method based on multi-scale feature fusion network

Abstract

The invention provides an unmanned aerial vehicle radio frequency signal detection and identification method based on a multi-scale feature fusion network. The recognition rate of the unmanned aerial vehicle under the small sample condition is mainly improved; according to the method, a 1D-ResAlex-FusionNet network is adopted, and the method comprises the following steps of: establishing a network; the 1D-ResAlex-FusionNet network comprises a CNN (Content Node Network)-ResNet module, an AlexNet-1D module and a feature fusion and classification module; the 1D-ResAlex FusionNet network is used for extracting a local feature vector through a CNN (Content Node Network)-ResNet module; the 1D-ResAlex-FusionNet network is used for extracting a multi-scale feature vector through an Alexne-1D module; the input of the feature fusion and classification module is a local feature vector and a multi-scale feature vector; the feature fusion and classification module splices and fuses the local feature vectors and the multi-scale feature vectors, classifies the fused feature vectors, and outputs a classification result of the radio frequency signals of the unmanned aerial vehicle; according to the method, the dependence on a single feature is reduced, the feature redundancy in a fusion structure is higher, the overfitting problem caused by a small sample can be effectively relieved, the influence of the number of samples under the condition of the small sample is smaller, and a more accurate detection result is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of detection and identification of UAV radio frequency signals, and specifically relates to a UAV radio frequency signal detection and identification method based on a multi-scale feature fusion network. Background Art

[0002] In modern society, drones are widely used in various fields. They are used in daily life to prevent and intervene in natural disasters, accelerate the mechanization of agriculture, and provide live broadcasts of sporting events and film shoots. Drone technology is also maturing rapidly. To ensure the healthy development of the drone industry, it is urgent to utilize modern communication and data processing technologies to achieve safe control of drones.

[0003] To achieve safe control of drones, the primary task is to effectively detect and identify their communication signals. Compared to other technical means, radio monitoring methods exhibit significant advantages, including low energy consumption, strong concealment, and the ability to obtain richer flight parameter information, thus possessing outstanding technical performance. This technology is mainly based on detecting the communication signals between the drone and its remote control device to achieve functions such as identification, positioning, and tracking of target drones. Research has shown that the use of a neural network architecture for radio signal detection can effectively improve the detection accuracy and classification capabilities of drone signals in complex electromagnetic environments. Currently, the method of directly inputting the raw communication signal into the neural network for training after basic preprocessing has a low effective detection and recognition rate for drones, especially when the available training sample is relatively small. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies and improve the recognition rate of drones in small sample sizes, the present invention provides a method for detecting and identifying drone radio frequency signals based on a multi-scale feature fusion network. By combining the CNN-ResNet and AlexNet-1D architectures, the present invention constructs a powerful feature extraction and fusion model that automatically extracts features from complex signals. By leveraging the respective strengths of the two networks, the model extracts more comprehensive features, enhancing the adaptability of drone signal detection for small sample sizes in practical applications and improving recognition accuracy and reliability in real-world environments.

[0005] The method for detecting and identifying UAV radio frequency signals based on a multi-scale feature fusion network includes the following steps:

[0006] Step 1: Preprocess the UAV RF signal data to obtain a multi-channel output signal; divide the multi-channel output signal into a training set, a validation set, and a test set;

[0007] Step 2: Build 1D-ResAlex-FusionNet network;

[0008] The input of the 1D-ResAlex-FusionNet network is a multi-channel output signal; the output of the 1D-ResAlex-FusionNet network is a classification result of the drone radio frequency signal; the 1D-ResAlex-FusionNet network includes a CNN-ResNet module, an AlexNet-1D module and a feature fusion and classification module; the 1D-ResAlex-FusionNet network extracts local feature vectors through the CNN-ResNet module; the 1D-ResAlex-FusionNet network extracts multi-scale feature vectors through the Alexnet-1D module; the input of the feature fusion and classification module is a local feature vector and a multi-scale feature vector; the feature fusion and classification module splices and fuses the local feature vector and the multi-scale feature vector, then classifies the fused feature vector, and outputs the classification result of the drone radio frequency signal;

[0009] Step 3: Use the training set to train the 1D-ResAlex-FusionNet network to obtain the optimal model of the 1D-ResAlex-FusionNet network;

[0010] Use the training set to train the 1D-ResAlex-FusionNet network, and use the validation set for validation at the end of each training cycle epoch; for each sample in the validation set, the 1D-ResAlex-FusionNet network outputs the classification result, compares the classification result of the sample in the validation set with the actual category of the sample, and counts the ratio of the number of correct classification results of samples in the validation set to the total number of samples in the validation set; the ratio of the number of correct classification results of samples in the validation set to the total number of samples in the validation set is the accuracy of the 1D-ResAlex-FusionNet network; reduce the learning rate of the 1D-ResAlex-FusionNet network when the accuracy stops improving; if the accuracy does not improve within the set training cycle after reducing the learning rate, stop training, and save the 1D-ResAlex-FusionNet network at this time as the optimal model;

[0011] Step 4: Use the optimal model of the 1D-ResAlex-FusionNet network obtained in step 3 to detect drone signals.

[0012] Furthermore, the method for preprocessing the drone RF signal data is a SMOTE algorithm; the SMOTE algorithm divides the data according to the specified number of channels and channel width to obtain a multi-channel output signal.

[0013] Furthermore, the CNN-ResNet module includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a first residual connection layer and a first flattening layer; the input of the CNN-ResNet module is a multi-channel output signal; the output of the CNN-ResNet module is a local feature vector; the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer and the first residual connection layer are connected in sequence; the output result of the second pooling layer and the output result of the first residual connection layer are directly added and input into the first flattening layer;

[0014] The first convolution layer includes sixty-four convolution kernels; the activation function of the first convolution layer is the ReLU activation function; the second convolution layer includes one hundred and twenty-eight convolution kernels; the sizes of the convolution kernels of the first convolution layer and the second convolution layer are both 3; the stride of the first pooling layer and the stride of the second pooling layer are both 2; the first residual connection layer includes the third convolution layer and the fourth convolution layer; the number of convolution kernels in the third convolution layer is 128; the number of convolution kernels in the fourth convolution layer is 256; the third convolution layer and the fourth convolution layer both use the ReLU activation function.

[0015] Furthermore, the input of the AlexNet-1D module is a multi-channel output signal; the output of the AlexNet-1D module is a multi-scale feature vector; the AlexNet-1D module includes a fifth convolutional layer, a third pooling layer, a sixth convolutional layer, a fourth pooling layer, a seventh convolutional layer, an eighth convolutional layer; a ninth convolutional layer, a fifth pooling layer, and a second flattening layer; the fifth convolutional layer, the third pooling layer, the sixth convolutional layer, the fourth pooling layer, the seventh convolutional layer, the eighth convolutional layer; the ninth convolutional layer, the fifth pooling layer, and the second flattening layer are connected in sequence; the fifth convolutional layer includes ninety-six convolution kernels; the size of the convolution kernel of the fifth convolutional layer is 11; The stride of the convolution kernel of the fifth convolution layer is 1; the activation function of the fifth convolution layer is the ReLU activation function; the sixth convolution layer includes 256 convolution kernels, and the convolution kernel size of the sixth convolution layer is 5; the seventh convolution layer includes 384 convolution kernels; the convolution kernel size of the seventh convolution layer is 3; the structure of the seventh convolution layer is the same as that of the eighth convolution layer; the ninth convolution layer includes 256 convolution kernels; the convolution kernel size of the ninth convolution layer is 3; the size of the third pooling layer is 2; the stride is 2; the structure of the fourth pooling layer is the same as that of the third pooling layer; the size of the fifth pooling layer is 2; the sixth, seventh and eighth convolution layers all use the ReLU activation function.

[0016] Furthermore, the feature fusion and classification module includes a splicing layer, a first fully connected layer, a dropout layer, and a second fully connected layer; the input of the feature fusion and classification module is a local feature vector and a multi-scale feature vector; the output of the feature fusion and classification module is a classification result of the UAV radio frequency signal; the splicing layer, the first fully connected layer, the dropout layer, and the second fully connected layer are connected in sequence;

[0017] The splicing layer splices the multi-scale feature vector and the local feature vector to obtain a fused feature vector; the first fully connected layer includes 1024 neurons; the activation function of the first fully connected layer is the ReLU activation function; the dropout rate of the Dropout layer is 0.5; the number of neurons in the second fully connected layer is the same as the number of signal types of drones; the number of signal types of drones is the number of categories of drones in the selected dataset; the activation function of the second fully connected layer is the Softmax activation function.

[0018] Furthermore, the loss function of the 1D-ResAlex-FusionNet network is a classification cross entropy loss function.

[0019] Furthermore, the optimizer of the 1D-ResAlex-FusionNet network is the Adam algorithm.

[0020] Furthermore, the initial learning rate of the Adam algorithm is set to 0.0005.

[0021] Furthermore, the Adam algorithm sets an "early stopping mechanism" and a learning rate decay mechanism.

[0022] The beneficial effects of the present invention are:

[0023] This paper combines two classic convolutional neural network architectures, CNN-ResNet and AlexNet-1D. The CNN-ResNet network solves the vanishing gradient problem of deep networks through residual connections, can extract deep abstract features, and uses a 3×3 convolution kernel that is more conducive to capturing local temporal patterns. AlexNet-1D uses convolution kernels of different sizes, such as 11x5x3, and is good at capturing multi-scale temporal features. Wider convolution kernels (such as 11) can identify long-period patterns, while small convolution kernels (such as 3) focus on details. By fusing the two types of features, the model can simultaneously utilize local details and global context, enhancing its ability to represent complex temporal patterns. The dual-branch structure forces the model to learn features from different perspectives. ResNet focuses on signal change trends, while AlexNet emphasizes multi-scale fluctuations. Reducing dependence on a single feature, the fusion structure has higher feature redundancy, can effectively alleviate the overfitting problem caused by small samples, is less affected by the number of samples in the case of small samples, has more accurate detection results, and has stronger adaptability in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a structural diagram of a network model of a UAV radio frequency signal detection method based on a multi-scale feature fusion network according to the present invention;

[0025] Figure 2 This is a diagram of the multi-channel 1DCNN-ResNet branch network structure used in the present invention;

[0026] Figure 3 This is a diagram showing the training results of a UAV radio frequency signal detection and recognition network based on a multi-scale feature fusion network in Example 1 of the present invention;

[0027] Figure 4 1 is a diagram showing the training results of a UAV radio frequency signal detection and recognition network based on a multi-scale feature fusion network in Example 1 of the present invention; wherein, (a) is the training result of the comparison method using 10% data as the training set; (b) is the training result of the present invention; (c) is the training result of the comparison method using 20% ​​data as the training set; and (d) is the training result of the present invention. DETAILED DESCRIPTION

[0028] The steps of the present invention are:

[0029] Step 1: Collect drone RF signal data, perform data enhancement on the drone RF signal data, divide it according to the specified number of channels and channel width, and obtain a multi-channel output signal; divide the multi-channel output signal into a training set, a validation set, and a test set;

[0030] The data enhancement method is the SMOTE algorithm; the SMOTE algorithm is the Synthetic Minority Over-sampling Technique; the data enhancement is performed on the acquired drone radio frequency signal data to solve the problem of class imbalance;

[0031] Simulate the distribution of actual signals in different frequency bands and convert the collected drone RF signal data into multi-channel input signals to better utilize convolutional neural networks to extract local features of the spectrum;

[0032] The multi-channel output signal is divided into training set, validation set and test set; the ratio of training set, validation set and test set is 8:1:1;

[0033] Step 2: Build 1D-ResAlex-FusionNet network;

[0034] The input of the 1D-ResAlex-FusionNet network is the multi-channel output signal; the output of the 1D-ResAlex-FusionNet network is the classification result of the drone RF signal;

[0035] The 1D-ResAlex-FusionNet network includes the CNN-ResNet module, the AlexNet-1D module, and the feature fusion and classification module;

[0036] The 1D-ResAlex-FusionNet network extracts local feature vectors through the CNN-ResNet module. The 1D-ResAlex-FusionNet network extracts multi-scale feature vectors through the Alexnet-1D module, then concatenates and fuses the local feature vectors with the multi-scale feature vectors to obtain a fused feature vector. Finally, after classification, the classification result of the drone's RF signal is output.

[0037] The CNN-ResNet module includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a first residual connection layer, and a first flattening layer; the input of the CNN-ResNet module is a multi-channel output signal; the output of the CNN-ResNet module is a local feature vector; the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, and the first residual connection layer are connected in sequence, and the output result of the second pooling layer and the output result of the first residual connection layer are directly added and input into the first flattening layer;

[0038] The first convolutional layer includes sixty-four convolution kernels; the activation function of the first convolutional layer is the ReLU activation function; the second convolutional layer includes one hundred and twenty-eight convolution kernels; the size of the convolution kernel of the first convolutional layer and the convolution kernel of the second convolutional layer are both 3; the stride of the first pooling layer and the stride of the second pooling layer are both 2; the first residual connection layer includes the third convolutional layer and the fourth convolutional layer; the number of convolution kernels in the third convolutional layer is 128; the number of convolution kernels in the fourth convolutional layer is 256; the third convolutional layer and the fourth convolutional layer both use the ReLU activation function;

[0039] The input of the AlexNet-1D module is a multi-channel output signal; the output of the AlexNet-1D module is a multi-scale feature vector;

[0040] The AlexNet-1D module includes a fifth convolutional layer, a third pooling layer, a sixth convolutional layer, a fourth pooling layer, a seventh convolutional layer, an eighth convolutional layer; a ninth convolutional layer, a fifth pooling layer, and a second flattening layer;

[0041] The fifth convolutional layer, the third pooling layer, the sixth convolutional layer, the fourth pooling layer, the seventh convolutional layer, the eighth convolutional layer; the ninth convolutional layer, the fifth pooling layer, and the second flattening layer are connected in sequence;

[0042] The fifth convolutional layer includes ninety-six convolution kernels; the size of the convolution kernel of the fifth convolutional layer is 11; the step size of the convolution kernel of the fifth convolutional layer is 1; the activation function of the fifth convolutional layer is the ReLU activation function;

[0043] The sixth convolutional layer includes 256 convolution kernels, and the convolution kernel size of the sixth convolutional layer is 5;

[0044] The seventh convolutional layer includes 384 convolution kernels; the convolution kernel size of the seventh convolutional layer is 3; the structure of the seventh convolutional layer is the same as that of the eighth convolutional layer;

[0045] The ninth convolutional layer includes 256 convolution kernels; the size of the convolution kernel of the ninth convolutional layer is 3;

[0046] The size of the third pooling layer is 2; the stride is 2;

[0047] The structure of the fourth pooling layer is the same as that of the third pooling layer;

[0048] The size of the fifth pooling layer is 2;

[0049] The sixth, seventh, and eighth convolutional layers all use the ReLU activation function;

[0050] The feature fusion and classification module includes a splicing layer, a first fully connected layer, a dropout layer, and a second fully connected layer; the input of the feature fusion and classification module is a local feature vector and a multi-scale feature vector; the output of the feature fusion and classification module is a classification result of the UAV radio frequency signal; the splicing layer, the first fully connected layer, the dropout layer, and the second fully connected layer are connected in sequence;

[0051] The concatenation layer concatenates the multi-scale feature vector and the local feature vector to obtain a fused feature vector; that is, in the area of ​​matching size on the concatenation axis, the corresponding slice of one feature vector is copied and placed at the end of the other feature vector; the first fully connected layer includes 1024 neurons; the activation function of the first fully connected layer is the ReLU activation function; the dropout rate of the Dropout layer is 0.5; the purpose of the Dropout layer is to prevent overfitting; the number of neurons in the second fully connected layer is the same as the number of drone signal types; the number of drone signal types is the number of drone categories in the selected dataset; the activation function of the second fully connected layer is the Softmax activation function;

[0052] Step 3: Construct the optimizer and loss function of the 1D-ResAlex-FusionNet network.

[0053] The loss function of the 1D-ResAlex-FusionNet network is the categorical cross entropy loss function, which is suitable for multi-class classification tasks and effectively measures the difference between the predicted value and the true label;

[0054] The optimizer of the 1D-ResAlex-FusionNet network is the Adam algorithm. The initial learning rate of the Adam algorithm is set to 0.0005, and the learning rate is dynamically adjusted by monitoring the accuracy of the validation set;

[0055] The Adam algorithm sets up an "early stopping mechanism" and a learning rate decay mechanism, uses the validation set and the "early stopping mechanism" to determine whether the model has converged, and uses learning rate decay to help the 1D-ResAlex-FusionNet network escape the local optimal solution;

[0056] Step 4: Train the 1D-ResAlex-FusionNet network;

[0057] The 1D-ResAlex-FusionNet network is trained using the training set, and at the end of each training cycle epoch, it is verified using the validation set. For each sample in the validation set, the 1D-ResAlex-FusionNet network outputs the classification result, compares the classification result of the sample in the validation set with the actual category of the sample, and counts the ratio of the number of correct classification results of samples in the validation set to the total number of samples in the validation set. The ratio of the number of correct classification results of samples in the validation set to the total number of samples in the validation set is the accuracy of the 1D-ResAlex-FusionNet network. When the accuracy stops improving, the learning rate of the 1D-ResAlex-FusionNet network is reduced. If the accuracy does not improve within the set training cycle after reducing the learning rate, the training is stopped, and the 1D-ResAlex-FusionNet network at this time is saved as the optimal model.

[0058] The present invention will be further described below with reference to the accompanying drawings and examples.

[0059] Example 1: The public drone dataset DroneRF is selected as the dataset for this example. The DroneRF dataset contains RF signals collected from three types of drones (Bebop, AR, and phantom) in five functional modes (off, powered on and connected, hovering, flying, and video recording). Specifically, the Bebop drone and the AR drone are tested in all five modes respectively. In addition, the phantom drone is only tested in the off and powered on and connected modes. The RF receiver is responsible for capturing the transmission signal between the drone and the controller, and its sampling bandwidth is 40MHz. Three different levels of detection and identification are performed on the drone existence (2 categories), drone model (4 categories), and drone status (10 categories). The specific implementation steps are as follows:

[0060] Step 1: Preprocess the dataset. Normalize the feature data to distribute them on the same scale. SMOTE upsampling is performed on the dataset to address sample imbalance in the original dataset. Channelization is used to convert the one-dimensional spectrum data into a multi-channel format (8 channels) to simulate the characteristics of signals in different frequency bands and enhance the model's ability to extract local spectral features. The dataset is evenly divided into a training set (80%), a test set (10%), and a validation set (10%) based on the labels.

[0061] Step 2: Design a 1D-ResAlex-FusionNet network to detect and identify drone RF signals, including individual and state recognition. This network takes processed multi-channel RF signal data as input, extracts feature vectors using the CNN-ResNet network and the Alexnet-1D network, then concatenates and fuses the feature vectors from the two networks to produce a fused feature vector. Finally, the classification layer outputs the classification results for the drone RF signal. The specific network construction is as follows:

[0062] Step 2.1: CNN-ResNet Construction: The input layer receives preprocessed RF signal data. First, a convolutional layer performs feature extraction. The first convolution operation uses 64 3×3 convolution kernels with a ReLU activation function. A 2×2 max pooling layer further extracts key signal features. A residual connection then passes the input directly to the output, enhancing the network's feature extraction capabilities. The residual block consists of two layers of convolution operations, with 128 and 256 kernels, respectively, using the ReLU activation function. Finally, a global average pooling layer converts the feature map into a one-dimensional feature vector.

[0063] Step 2.2: AlexNet-1D Construction: The input layer receives the same preprocessed RF signal data. The input data undergoes multiple convolution and pooling operations to extract progressively deeper features. The first convolution layer uses 96 11×11 kernels with a stride of 4 and a ReLU activation function. The subsequent max pooling layer uses a 3×3 pooling window with a stride of 2. The second convolution layer uses 256 5×5 kernels with the same padding and a ReLU activation function. This is followed by the same max pooling layer as before. Subsequent convolution layers use 384, 384, and 256 3×3 kernels with the same padding, respectively, and all use ReLU activation functions. Finally, a global average pooling layer converts the feature map into a one-dimensional feature vector.

[0064] Step 2.3: Feature Fusion and Classification: Concatenate and fuse the feature vectors extracted by CNN-ResNet and AlexNet-1D to generate a fused feature vector. Add a fully connected layer after the fused feature vector, using the ReLU activation function, to further process the fused features. Add a Dropout layer to prevent overfitting and improve the model's generalization. The final layer is the classification layer, using the softmax activation function, to output the classification results of the drone's RF signal.

[0065] Step 3: Set up the optimizer and loss function. The model uses the categorical cross entropy loss function, which is suitable for multi-class classification tasks and effectively measures the difference between predicted values ​​and true labels. The optimizer uses the Adam algorithm, with an initial learning rate set to 0.0005. The learning rate is dynamically adjusted by monitoring the accuracy of the validation set. An early stopping mechanism and learning rate decay are implemented, using the validation set and the early stopping mechanism to determine model convergence. Learning rate decay is used to help the model escape from local optima.

[0066] Step 4: Model Training. In this stage, we first set the training parameters. The maximum number of training epochs was set to 300, the batch size was set to 128, and the initial learning rate was set to 0.0005. The patience value for the early stopping mechanism was set to 50; if the validation set performance did not improve within 50 consecutive epochs, training would be stopped early. The patience value for the learning rate decay was set to 20; if the validation set performance did not improve within 20 consecutive epochs, the learning rate would be halved. The processed data was then fed into the network for training to obtain the model with the highest accuracy.

[0067] To demonstrate the superiority of the present invention, simulations were conducted to compare it with the Multi-Channel 1D CNN network used in the paper [Deep Learning for RF-Based Drone Detection and Identification: A Multi-Channel 1-D Convolutional Neural Networks Approach]. Similarly, the maximum number of training rounds was set to 300, the batch size was set to 128, and the initial learning rate was set to 0.0005. Classification accuracy was selected as the evaluation metric, defined as the ratio of the number of correctly classified test samples to the total number of test samples. The accuracy performance of the present invention and the comparison scheme on various classification tasks is shown in Table 1.

[0068] Table 1 Accuracy of this method and the comparison method on different classification tasks

[0069]

[0070] As shown in Table 1, with identical parameter settings, the present invention achieves higher accuracy than the comparison solution for drone detection and type recognition, as well as drone detection, type, and state recognition. In particular, the accuracy for the most complex ten-class task increased by 4% compared to the comparison solution, achieving this improvement without introducing more complex and time-consuming feature extraction.

[0071] Example 2: Testing the effect of the present invention on a small sample data set.

[0072] The public drone dataset DroneRF was selected as the dataset for this example. The DroneRF dataset contains RF signals collected from three types of drones (Bebop, AR, and phantom) in five functional modes (off, powered on and connected, hovering, flying, and video recording). Specifically, the Bebop drone and the AR drone were tested in all five modes. In addition, the phantom drone was only tested in the off and powered on and connected modes. The RF receiver is responsible for capturing the transmission signal between the drone and the controller, and its sampling bandwidth is 40MHz. Three different levels of detection and recognition are performed on the drone's presence (2 categories), drone model (4 categories), and drone status (10 categories). The specific implementation steps are as follows:

[0073] Step 1: Preprocess the dataset. Normalize the feature data to distribute them on the same scale. SMOTE upsampling is performed on the dataset to address sample imbalance in the original dataset. Channelization is used to convert the one-dimensional spectrum data into a multi-channel format (8 channels), simulating the characteristics of signals in different frequency bands and enhancing the model's ability to extract local spectral features. The dataset is evenly divided into a training set (80%), a test set (10%), and a validation set (10%) based on the labels. A small portion of the training set (e.g., 10%) is then allocated as the small sample dataset used in this example.

[0074] Step 2: Design a 1D-ResAlex-FusionNet network to detect and identify drone RF signals, including individual and state recognition. This network takes processed multi-channel RF signal data as input, extracts feature vectors using the CNN-ResNet network and the Alexnet-1D network, then concatenates and fuses the feature vectors from the two networks to produce a fused feature vector. Finally, the classification layer outputs the classification results for the drone RF signal. The specific network construction is as follows:

[0075] Step 2.1: CNN-ResNet Construction: The input layer receives preprocessed RF signal data. First, a convolutional layer performs feature extraction. The first convolution operation uses 64 3×3 convolution kernels with a ReLU activation function. A 2×2 max pooling layer further extracts key signal features. A residual connection then passes the input directly to the output, enhancing the network's feature extraction capabilities. The residual block consists of two layers of convolution operations, with 128 and 256 kernels, respectively, using the ReLU activation function. Finally, a global average pooling layer converts the feature map into a one-dimensional feature vector.

[0076] Step 2.2: AlexNet-1D Construction: The input layer receives the same preprocessed RF signal data. The input data undergoes multiple convolution and pooling operations to extract progressively deeper features. The first convolution layer uses 96 11×11 kernels with a stride of 4 and a ReLU activation function. The subsequent max pooling layer uses a 3×3 pooling window with a stride of 2. The second convolution layer uses 256 5×5 kernels with the same padding and a ReLU activation function. This is followed by the same max pooling layer as before. Subsequent convolution layers use 384, 384, and 256 3×3 kernels with the same padding, respectively, and all use ReLU activation functions. Finally, a global average pooling layer converts the feature map into a one-dimensional feature vector.

[0077] Step 2.3: Feature Fusion and Classification: Concatenate and fuse the feature vectors extracted by CNN-ResNet and AlexNet-1D to generate a fused feature vector. Add a fully connected layer after the fused feature vector, using the ReLU activation function, to further process the fused features. Add a Dropout layer to prevent overfitting and improve the model's generalization. The final layer is the classification layer, using the softmax activation function, to output the classification results of the drone's RF signal.

[0078] Step 3: Set up the optimizer and loss function. The model uses the categorical cross entropy loss function, which is suitable for multi-class classification tasks and effectively measures the difference between predicted values ​​and true labels. The optimizer uses the Adam algorithm, with an initial learning rate set to 0.0005. The learning rate is dynamically adjusted by monitoring the accuracy of the validation set. An early stopping mechanism and learning rate decay are implemented, using the validation set and the early stopping mechanism to determine model convergence. Learning rate decay is used to help the model escape from local optima.

[0079] Step 4: Model Training. In this stage, we first set the training parameters. The maximum number of training epochs was set to 300, the batch size was set to 128, and the initial learning rate was set to 0.0005. The patience value for the early stopping mechanism was set to 50; if the validation set performance did not improve within 50 consecutive epochs, training would be stopped early. The patience value for the learning rate decay was set to 20; if the validation set performance did not improve within 20 consecutive epochs, the learning rate would be halved. The processed data was then fed into the network for training to obtain the model with the highest accuracy.

[0080] To verify the superiority of this invention, simulations were conducted in comparison with the Multi-Channel 1D CNN network used in the paper [Deep Learning for RF-Based Drone Detection and Identification: A Multi-Channel 1-D Convolutional Neural Networks Approach]. Similarly, the maximum number of training rounds was set to 300, the batch size was set to 128, and the initial learning rate was set to 0.0005. The evaluation metric chosen was classification accuracy, defined as the ratio of the number of correctly classified test samples to the total number of test samples.

[0081] The final results of selecting 10% and 20% of the original training set data as small sample training sets in the drone detection, type and state recognition (10 categories) tasks are shown in Table 2.

[0082] Table 2 Classification results of this method and the comparison method under small sample conditions

[0083]

[0084] It can be seen from Table 2 that when 10% of the data is used as the training set, the accuracy of the comparative experiment is only 50%, while the accuracy of the present invention reaches 76%, which is 52% higher than that of the comparative experiment; when 20% of the data is used as the training set, the accuracy of the comparative experiment is only 74.7%, while the accuracy of the present invention reaches 81.5%, which is 9% higher than that of the comparative experiment; under the same parameter settings, the classification accuracy of the present invention in the case of small samples is greatly improved compared with the network in the comparative experiment, and the effect is more prominent on a smaller training set.

Claims

1. A method for detecting and identifying UAV radio frequency signals based on a multi-scale feature fusion network, characterized in that: The steps include: Step 1: Preprocess the UAV RF signal data to obtain a multi-channel output signal; divide the multi-channel output signal into a training set, a validation set, and a test set; Step 2: Build 1D-ResAlex-FusionNet network; The input of the 1D-ResAlex-FusionNet network is a multi-channel output signal; the output of the 1D-ResAlex-FusionNet network is a classification result of the drone radio frequency signal; the 1D-ResAlex-FusionNet network includes a CNN-ResNet module, an AlexNet-1D module and a feature fusion and classification module; the 1D-ResAlex-FusionNet network extracts local feature vectors through the CNN-ResNet module; the 1D-ResAlex-FusionNet network extracts multi-scale feature vectors through the Alexnet-1D module; the input of the feature fusion and classification module is a local feature vector and a multi-scale feature vector; the feature fusion and classification module splices and fuses the local feature vector and the multi-scale feature vector, then classifies the fused feature vector, and outputs the classification result of the drone radio frequency signal; Step 3: Use the training set to train the 1D-ResAlex-FusionNet network to obtain the optimal model of the 1D-ResAlex-FusionNet network; Use the training set to train the 1D-ResAlex-FusionNet network, and use the validation set for validation at the end of each training cycle epoch; for each sample in the validation set, the 1D-ResAlex-FusionNet network outputs the classification result, compares the classification result of the sample in the validation set with the actual category of the sample, and counts the ratio of the number of correct classification results of samples in the validation set to the total number of samples in the validation set; the ratio of the number of correct classification results of samples in the validation set to the total number of samples in the validation set is the accuracy of the 1D-ResAlex-FusionNet network; reduce the learning rate of the 1D-ResAlex-FusionNet network when the accuracy stops improving; if the accuracy does not improve within the set training cycle after reducing the learning rate, stop training, and save the 1D-ResAlex-FusionNet network at this time as the optimal model; Step 4: Use the optimal model of the 1D-ResAlex-FusionNet network obtained in step 3 to detect drone signals.

2. The method for detecting and identifying UAV radio frequency signals based on a multi-scale feature fusion network according to claim 1 is characterized in that: The method for preprocessing the drone RF signal data is the SMOTE algorithm; the SMOTE algorithm divides the data according to the specified number of channels and channel width to obtain a multi-channel output signal.

3. The method for detecting and identifying UAV radio frequency signals based on a multi-scale feature fusion network according to claim 1 is characterized in that: The CNN-ResNet module includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a first residual connection layer and a first flattening layer; the input of the CNN-ResNet module is a multi-channel output signal; the output of the CNN-ResNet module is a local feature vector; the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer and the first residual connection layer are connected in sequence; the output result of the second pooling layer and the output result of the first residual connection layer are directly added and input into the first flattening layer; The first convolution layer includes sixty-four convolution kernels; the activation function of the first convolution layer is the ReLU activation function; the second convolution layer includes one hundred and twenty-eight convolution kernels; the sizes of the convolution kernels of the first convolution layer and the second convolution layer are both 3; the stride of the first pooling layer and the stride of the second pooling layer are both 2; the first residual connection layer includes the third convolution layer and the fourth convolution layer; the number of convolution kernels in the third convolution layer is 128; the number of convolution kernels in the fourth convolution layer is 256; the third convolution layer and the fourth convolution layer both use the ReLU activation function.

4. The method for detecting and identifying UAV radio frequency signals based on a multi-scale feature fusion network according to claim 1 is characterized in that: The input of the AlexNet-1D module is a multi-channel output signal; the output of the AlexNet-1D module is a multi-scale feature vector; the AlexNet-1D module includes a fifth convolutional layer, a third pooling layer, a sixth convolutional layer, a fourth pooling layer, a seventh convolutional layer, an eighth convolutional layer; a ninth convolutional layer, a fifth pooling layer and a second flattening layer; the fifth convolutional layer, the third pooling layer, the sixth convolutional layer, the fourth pooling layer, the seventh convolutional layer, the eighth convolutional layer; the ninth convolutional layer, the fifth pooling layer and the second flattening layer are connected in sequence; the fifth convolutional layer includes ninety-six convolution kernels; the size of the convolution kernel of the fifth convolutional layer is 1 ... fifth convolutional layer includes ninety-six convolution kernels; the fifth convolutional layer includes ninety-six convolution kernels; the fifth convolutional layer includes ninety-six convolution kernels; the fifth convolutional layer includes ninety-six convolution kernels; the fifth convolutional layer includes ninety-six convolution kernels; the fifth convolutional layer includes ninety-six convolution kernels; the fifth convolutional layer includes ninety-six convolution kernels; the fifth convolutional layer includes ninety-six convolution kernels; the fifth convolutional layer includes The stride of the convolution kernel of the convolution layer is 1; the activation function of the fifth convolution layer is the ReLU activation function; the sixth convolution layer includes 256 convolution kernels, and the convolution kernel size of the sixth convolution layer is 5; the seventh convolution layer includes 384 convolution kernels; the convolution kernel size of the seventh convolution layer is 3; the structure of the seventh convolution layer is the same as that of the eighth convolution layer; the ninth convolution layer includes 256 convolution kernels; the convolution kernel size of the ninth convolution layer is 3; the size of the third pooling layer is 2; the stride is 2; the structure of the fourth pooling layer is the same as that of the third pooling layer; the size of the fifth pooling layer is 2; the sixth, seventh and eighth convolution layers all use the ReLU activation function.

5. The method for detecting and identifying UAV radio frequency signals based on a multi-scale feature fusion network according to claim 1, characterized in that: The feature fusion and classification module includes a splicing layer, a first fully connected layer, a dropout layer, and a second fully connected layer; the input of the feature fusion and classification module is a local feature vector and a multi-scale feature vector; the output of the feature fusion and classification module is a classification result of the UAV radio frequency signal; the splicing layer, the first fully connected layer, the dropout layer, and the second fully connected layer are connected in sequence; The splicing layer splices the multi-scale feature vector and the local feature vector to obtain a fused feature vector; the first fully connected layer includes 1024 neurons; the activation function of the first fully connected layer is the ReLU activation function; the dropout rate of the Dropout layer is 0.5; the number of neurons in the second fully connected layer is the same as the number of signal types of drones; the number of signal types of drones is the number of categories of drones in the selected dataset; the activation function of the second fully connected layer is the Softmax activation function.

6. The method for detecting and identifying UAV radio frequency signals based on a multi-scale feature fusion network according to claim 1, characterized in that: The loss function of the 1D-ResAlex-FusionNet network is the classification cross entropy loss function.

7. The method for detecting and identifying UAV radio frequency signals based on a multi-scale feature fusion network according to claim 1, characterized in that: The optimizer of the 1D-ResAlex-FusionNet network is the Adam algorithm.

8. The method for detecting and identifying UAV radio frequency signals based on a multi-scale feature fusion network according to claim 7, characterized in that: The initial learning rate of the Adam algorithm is set to 0.0005.

9. The method for detecting and identifying UAV radio frequency signals based on a multi-scale feature fusion network according to claim 8, characterized in that: The Adam algorithm sets an "early stopping mechanism" and a learning rate decay mechanism.

Images