Unmanned aerial vehicle identification method and device based on category-specific modeling, equipment and storage medium

Abstract

The application discloses a UAV identification method and device based on category-specific modeling, equipment and storage medium, relates to the pattern recognition technology field based on digital data processing, and the method comprises the steps of acquiring the radio frequency signal of a UAV, generating a time-frequency graph through short-time Fourier transform; inputting the time-frequency graph into a target residual network to extract a UAV feature vector; using a Gaussian mixture model constructed for each known class of UAV in the training stage to calculate the log-likelihood value of the feature vector under each model, and obtaining a preliminary discrimination result according to the maximum log-likelihood value and a preset adaptive threshold; for the sample preliminarily judged as an unknown class, calculating the minimum feature distance of the sample from all samples in the feature matrix of the known-class UAV samples, and if the distance is less than a preset recovery threshold, reclassifying the sample into the known-class UAV to which the nearest neighbor sample belongs. The application can accurately identify known-class UAVs and effectively distinguish unknown-class UAVs in an open environment.

Description

Technical Field

[0001] This application relates to the field of pattern recognition technology based on digital data processing, and in particular to a method, apparatus, device and storage medium for drone recognition based on category-specific modeling. Background Technology

[0002] With the widespread application of drones in agriculture, surveying, logistics, and emergency rescue, illegal uses such as unauthorized and excessive drone flights are becoming increasingly frequent, seriously threatening airspace safety and public order. Against this backdrop, effective identification and legality verification of drones have become a crucial link in low-altitude security systems. Radio frequency signal analysis, due to its advantages such as long operating range, strong concealment, and ability to penetrate obstructions, is widely regarded as an effective drone detection method. Especially in open environments, the system not only needs to accurately identify known drone models but also needs to have the ability to reject unknown models not encountered during the training phase—that is, to achieve open-set identification. This places higher demands on ensuring the robustness and generalization capabilities of the monitoring system.

[0003] Currently, most mainstream drone identification methods are based on closed-set classification of time-frequency graphs of radio frequency signals using deep learning models. For example, some studies use convolutional neural networks such as residual networks to extract time-frequency features and combine them with cross-entropy loss functions for training; other works introduce attention mechanisms or improve loss functions to enhance feature discriminativeness. In addition, some open-set identification methods attempt to distinguish between known and unknown classes through OpenMax (Open Set Maximum Probability Method), adversarial learning, or multi-centroid modeling.

[0004] However, existing approaches still have significant shortcomings: on the one hand, most models only focus on intra-class aggregation in the feature space, neglecting inter-class separation, leading to blurred boundaries between different classes and affecting open-set discrimination performance; on the other hand, UAV radio frequency signals have complex multi-peak, non-uniform distribution characteristics, which traditional single Gaussian assumptions or simple classifiers cannot accurately model. Although some studies have attempted to introduce Gaussian mixture models, they are often not effectively combined with class-specific modeling and deep feature extraction, resulting in limited ability to reject unknown classes. Furthermore, under complex electromagnetic interference, the model's ability to perceive key time-frequency regions is insufficient, further restricting recognition accuracy and stability. Therefore, how to accurately identify known UAV classes and effectively distinguish unknown UAV classes in open environments has become an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this application is to provide a method, apparatus, device and storage medium for drone identification based on category-specific modeling, which aims to solve the technical problem of how to accurately identify known types of drones and effectively distinguish unknown types of drones in an open environment.

[0006] To achieve the above objectives, this application proposes a drone identification method based on category-specific modeling, the method comprising:

[0007] The radio frequency signals of the UAVs within the monitoring area are acquired, and the radio frequency signals are subjected to short-time Fourier transform to obtain a time-frequency diagram;

[0008] The time-frequency graph is input into the target residual network for feature extraction to obtain the UAV feature vector. The target residual network is an improvement on ResNet18.

[0009] The corresponding Gaussian mixture model constructed for each known class of UAV during the training phase is invoked, and the log-likelihood value of the UAV feature vector under each Gaussian mixture model is calculated.

[0010] A preliminary discrimination result is obtained based on the maximum value among the log-likelihood values ​​and a preset adaptive threshold, wherein the preset adaptive threshold is determined during the training phase;

[0011] The feature matrix of the known class drone samples obtained in the training phase and the preset recovery threshold are called, and the minimum feature distance between the drone feature vector corresponding to the unknown class drone in the preliminary discrimination result and all samples in the feature matrix is ​​calculated.

[0012] When the minimum feature distance is less than the preset recovery threshold, the unknown type of UAV is reclassified as the known type of UAV corresponding to the minimum feature distance.

[0013] In one embodiment, the target residual network includes convolutional layers, pooling layers, convolutional block attention modules, feature distribution normalization modules, L2 regularization modules, fully connected layers, and an output layer.

[0014] The step of inputting the time-frequency map into the target residual network for feature extraction to obtain the UAV feature vector includes:

[0015] The time-frequency image is input into the target residual network, and features are extracted from the time-frequency image through the convolutional layer to obtain local features;

[0016] The local features are downsampled using the pooling layer to obtain preliminary features;

[0017] The initial features are enhanced by the convolutional block attention module in both channel and spatial dimensions to obtain enhanced features.

[0018] The enhanced features are processed by the feature distribution standardization module to obtain standardized features.

[0019] The standardized features are regularized using the L2 regularization processing module to obtain regularized features;

[0020] The UAV feature vector is obtained by performing dimension mapping and feature integration on the regularized features through the fully connected layer, and then output through the output layer.

[0021] In one embodiment, the construction steps of the target residual network include:

[0022] The ResNet18 is used as the basic network architecture, which includes convolutional layers, pooling layers, fully connected layers, and an output layer.

[0023] A convolutional block attention module is embedded after the pooling layer, and a feature distribution normalization module and an L2 regularization processing module are embedded before the fully connected layer to obtain the initial residual network.

[0024] By combining the cross-entropy loss function with the contrast center loss function, an improved loss function is obtained.

[0025] The improved loss function is used as the loss calculation function of the initial residual network to obtain the target residual network.

[0026] In one embodiment, the training steps of the target residual network include:

[0027] Obtain a training dataset and a validation dataset, both of which include radio frequency signal samples and time-frequency plot samples converted from the radio frequency signal samples, wherein the radio frequency signal samples contain known class labels;

[0028] The time-frequency graph samples are divided into multiple sample batches according to a preset batch size;

[0029] The target residual network to be trained is used as the initial network, and each batch of samples is sequentially input into the initial network to obtain the prediction results.

[0030] Calculate the loss value between the prediction result and the corresponding known class label;

[0031] Based on the loss value, the parameters of the target residual network to be trained are adjusted using the backpropagation algorithm, and the process of inputting each batch of samples into the target residual network to be trained in sequence to obtain the prediction result is repeated until the number of iterations reaches the preset number of training rounds to obtain the reference network.

[0032] The recognition accuracy of the reference network is calculated based on the validation dataset;

[0033] When the recognition accuracy is less than the preset accuracy threshold, return to the step of sequentially inputting each batch of samples into the target residual network to be trained to obtain the prediction result;

[0034] When the recognition accuracy is greater than or equal to the preset accuracy threshold, the reference network is used as the target residual network after training.

[0035] In one embodiment, the steps for obtaining the feature matrix and the preset recycling threshold include:

[0036] During the training phase, training samples of all known types of UAVs are acquired;

[0037] The target residual network is used to extract features from each training sample to obtain a sample feature vector corresponding to each training sample.

[0038] The sample feature vectors are classified and organized according to the known class labels of the training samples, and all the sample feature vectors corresponding to the same known class label are combined to obtain the feature matrix of the known class UAV samples;

[0039] Calculate the pairwise Euclidean distance between any two sample feature vectors in the feature matrix to obtain a distance matrix, wherein the elements at the diagonal positions in the distance matrix are set to infinity;

[0040] The minimum distance between each sample feature vector and all other sample feature vectors is selected from the distance matrix to obtain the minimum distance set;

[0041] The 95th percentile of the minimum distance set is used as the preset recycling threshold.

[0042] In one embodiment, the step of acquiring the radio frequency signal of the UAV within the monitoring area and performing a short-time Fourier transform on the radio frequency signal to obtain a time-frequency diagram includes:

[0043] Acquire radio frequency signals of UAVs within the monitoring area, the radio frequency signals including uplink non-payload control signals and downlink image transmission signals;

[0044] The radio frequency signal is segmented into segments by a preset sliding window function to obtain multiple local signals;

[0045] Perform a Fourier transform on each of the local signals to obtain the frequency domain feature data corresponding to each local signal;

[0046] The power spectral density of the radio frequency signal is calculated based on all the frequency domain feature data, and a time-frequency diagram is generated based on the power spectral density.

[0047] In one embodiment, the step of obtaining a preliminary discrimination result based on the maximum value among the log-likelihood values ​​and a preset adaptive threshold includes:

[0048] When the maximum value in the log-likelihood values ​​is greater than or equal to a preset adaptive threshold, a preliminary judgment result is obtained that the drone is a known class drone, and the known class drone belongs to the known drone class corresponding to the Gaussian mixture model to which the maximum value belongs;

[0049] When the maximum value is less than the preset adaptive threshold, a preliminary judgment result is obtained that the drone is an unknown type of drone.

[0050] Furthermore, to achieve the above objectives, this application also proposes a drone identification device based on category-specific modeling, the device comprising:

[0051] The time-frequency diagram generation module is used to acquire the radio frequency signals of the UAV within the monitoring area and perform a short-time Fourier transform on the radio frequency signals to obtain a time-frequency diagram.

[0052] The feature extraction module is used to input the time-frequency map into the target residual network for feature extraction to obtain the UAV feature vector. The target residual network is an improvement on ResNet18.

[0053] The likelihood calculation module is used to call the corresponding Gaussian mixture model built for each known class of UAV during the training phase, and calculate the log-likelihood value of the UAV feature vector under each Gaussian mixture model.

[0054] The preliminary discrimination module is used to compare the maximum value in the log-likelihood values ​​with a preset adaptive threshold to obtain a preliminary discrimination result. The preset adaptive threshold is determined during the training phase.

[0055] The distance calculation module is used to call the feature matrix of the known class drone samples obtained in the training phase and the preset recovery threshold, and to calculate the minimum feature distance between the drone feature vector corresponding to the unknown class drone in the preliminary discrimination result and all samples in the feature matrix;

[0056] The secondary discrimination module is used to reclassify the unknown type drone as the known type drone corresponding to the minimum feature distance when the minimum feature distance is less than the preset recovery threshold.

[0057] Furthermore, to achieve the above objectives, this application also proposes a drone identification device based on category-specific modeling, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the drone identification method based on category-specific modeling as described above.

[0058] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the UAV identification method based on category-specific modeling as described above.

[0059] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the category-specific modeling-based UAV identification method described above.

[0060] One or more technical solutions proposed in this application have at least the following technical effects:

[0061] First, the radio frequency signals of drones within the monitoring area are acquired and subjected to a short-time Fourier transform to generate a time-frequency map that simultaneously characterizes the signal's temporal evolution and frequency components, providing a two-dimensional input rich in discriminative information for subsequent identification. Next, this time-frequency map is input into a target residual network for feature extraction, yielding a compact and discriminative drone feature vector. This network, based on an improvement of ResNet18, maintains deep feature extraction capabilities while enhancing its focus on key time-frequency regions and optimizing the feature space structure. Subsequently, Gaussian mixture models, constructed separately for each known class of drone during the training phase, are invoked to calculate the logarithmic similarity of the drone's feature vector under each model. The system calculates the minimum feature distance between the initial classification of a UAV and a preset adaptive threshold. The maximum log-likelihood value is compared with this threshold, which is dynamically determined based on the likelihood distribution of the training and validation data, improving the stability of open-set classification. To further reduce the misclassification of known samples as unknown classes, the system uses the feature matrix of known class UAV samples built during training and a preset retrieval threshold to calculate the minimum feature distance between the feature vector of the UAV initially classified as unknown and all known class samples. If this distance is less than the preset retrieval threshold, the UAV is reclassified as belonging to the known class of its nearest neighbor, effectively correcting the conservative rejection that might occur with the Gaussian mixture model. This application, through a multi-stage collaborative classification mechanism, can accurately identify known class UAVs and effectively distinguish unknown class UAVs in open environments. Attached Figure Description

[0062] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0063] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a flowchart illustrating an embodiment of the UAV identification method based on category-specific modeling in this application.

[0065] Figure 2 This is a schematic diagram of the target residual network provided in Embodiment 1 of the UAV recognition method based on category-specific modeling in this application;

[0066] Figure 3 This is a flowchart illustrating Embodiment 2 of the UAV identification method based on category-specific modeling in this application;

[0067] Figure 4 A schematic diagram of the processing flow of the convolutional block attention module in the UAV recognition method based on category-specific modeling provided in Embodiment 2 of this application;

[0068] Figure 5 This is a schematic diagram of the module structure of the drone identification device based on category-specific modeling according to an embodiment of this application;

[0069] Figure 6 This is a schematic diagram of the device structure of the hardware operating environment involved in the category-specific modeling-based drone identification method in this application embodiment.

[0070] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0071] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0072] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0073] It should be noted that the executing entity of this application embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or drone signal recognition system capable of performing the above functions. The following description uses a drone signal recognition system as an example to illustrate this embodiment and the subsequent embodiments.

[0074] Based on this, embodiments of this application provide a drone identification method based on category-specific modeling, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the drone identification method based on category-specific modeling in this application.

[0075] In this embodiment, the UAV identification method based on category-specific modeling includes steps S10 to S60:

[0076] Step S10: Acquire the radio frequency signal of the UAV in the monitoring area, and perform a short-time Fourier transform on the radio frequency signal to obtain a time-frequency diagram.

[0077] It should be noted that the monitoring area refers to a specific airspace or geographical area where radio frequency (RF) signal receiving equipment is deployed for the detection and identification of drones entering the area. RF signals are electromagnetic wave signals emitted by a drone during flight to communicate with its remote controller or ground station. They typically include flight control commands (uplink) and image transmission data (downlink), and their frequencies are generally located in unlicensed frequency bands (e.g., 2.4 GHz or 5.8 GHz). A time-frequency graph is a two-dimensional image representation obtained by performing a Short-Time Fourier Transform (STFT) on the RF signal. The horizontal axis represents time, and the vertical axis represents frequency. Pixel intensity reflects the energy distribution of the corresponding frequency at that moment, visually presenting the joint characteristics of the signal in both time and frequency dimensions.

[0078] As an example, the step of acquiring the radio frequency (RF) signal of a UAV within the monitoring area and performing a short-time Fourier transform on the RF signal to obtain a time-frequency graph includes: acquiring the RF signal of the UAV within the monitoring area, the RF signal including an uplink non-load control signal and a downlink image transmission signal; segmenting the RF signal segment by segment using a preset sliding window function to obtain multiple local signals; performing a Fourier transform on each local signal to obtain frequency domain feature data corresponding to each local signal; calculating the power spectral density of the RF signal based on all the frequency domain feature data, and generating a time-frequency graph based on the power spectral density.

[0079] Uplink non-payload control signals refer to radio frequency (RF) command signals sent from ground-based remote control equipment to the UAV to control its flight parameters such as attitude, heading, and altitude. These signals do not contain mission payload data (e.g., images or video) and are used solely for flight control. Downlink image transmission signals refer to the RF signals corresponding to real-time video or image data streams transmitted from the UAV to the ground receiver. They typically carry first-person view (FPV) images or other sensor information and are used for operator monitoring or mission execution. A preset sliding window function (e.g., a Hamming window) is a window function used in short-time Fourier transform to extract local segments of a signal. Its shape and length are preset and it slides along the time axis with a fixed step size to achieve piecewise stabilization of non-stationary signals. A local signal refers to a finite-length sub-signal extracted from the original RF signal using a sliding window function. This sub-signal is considered approximately stationary for subsequent frequency domain analysis. Frequency domain feature data refers to the complex sequence obtained after performing a Fourier transform on each local signal. The square of its modulus reflects the energy distribution of each frequency component within that local time period and is the basic unit for constructing time-frequency representations. Power spectral density refers to the signal power distribution per unit frequency. It is obtained by calculating the square of the amplitude of multiple local signals' frequency domain feature data and then normalizing and averaging it. It is used to characterize the energy intensity of radio frequency signals in the joint time and frequency dimensions and is the core basis for generating time-frequency maps.

[0080] First, the UAV signal identification system continuously collects aerial electromagnetic signals through radio frequency receiving equipment deployed within the monitoring area, and filters out radio frequency signals belonging to UAV communication. These signals include uplink non-payload control signals sent from the ground remote controller to the UAV and downlink image transmission signals transmitted back to the ground by the UAV. Second, the system uses a preset sliding window function to segment the radio frequency signal along the time axis from the start time, moving a fixed step each time, thereby dividing the entire non-stationary signal into multiple continuous and partially overlapping local signals to meet the requirements of local signal stationarity for subsequent frequency domain analysis. Then, the system performs a Discrete Fourier Transform (DFT) on each segmented local signal, calculates its complex spectrum at each frequency point, and further calculates the square of the amplitude to obtain the frequency domain feature data corresponding to each local signal. Finally, the system arranges the frequency domain feature data of all local signals in chronological order, calculates their power spectral density (i.e., average energy per unit frequency), and maps this power spectral density to a grayscale or pseudo-color image, with the horizontal axis representing time, the vertical axis representing frequency, and pixel brightness representing energy intensity, thereby generating a time-frequency map for subsequent identification.

[0081] Step S20: Input the time-frequency map into the target residual network for feature extraction to obtain the UAV feature vector. The target residual network is an improvement on ResNet18.

[0082] It should be noted that the target residual network refers to a deep convolutional neural network obtained by making specific improvements to the original ResNet18 network structure. While retaining the basic architecture of the residual blocks, it integrates additional modules (such as attention mechanisms) or adopts optimized loss functions to enhance the ability to extract discriminative features from the UAV time-frequency image. The UAV feature vector is a fixed-dimensional real-number vector output at the top layer of the target residual network after performing layer-by-layer convolution, pooling, and nonlinear transformations on the input time-frequency image. This vector condenses the high-level semantic information related to the UAV model in the input time-frequency image and is used for subsequent classification or modeling. ResNet18 refers to a residual network (Residual Network) proposed by Microsoft Research, containing 18 weight layers. It alleviates the gradient vanishing problem in deep network training by introducing cross-layer identity mappings (i.e., "residual connections") and is widely used in visual tasks such as image classification. It is the basic architecture of the target residual network in this embodiment.

[0083] Understandably, the UAV signal recognition system inputs the time-frequency map into the target residual network. The target residual network sequentially performs forward propagation calculations on the time-frequency map through its convolutional layers, pooling layers, and multiple residual blocks, extracting hierarchical feature representations from local to global layers, and outputs a fixed-dimensional real vector before the final fully connected layer of the network. This vector is the UAV feature vector.

[0084] As an example, the construction steps of the target residual network include: using ResNet18 as the basic network architecture, which includes convolutional layers, pooling layers, fully connected layers, and an output layer; embedding a convolutional block attention module after the pooling layer, and embedding a feature distribution normalization module and an L2 regularization processing module before the fully connected layer to obtain an initial residual network; combining the cross-entropy loss function with the contrast center loss function to obtain an improved loss function; and using the improved loss function as the loss calculation function of the initial residual network to obtain the target residual network.

[0085] The basic network architecture refers to the overall structure of a deep neural network built based on ResNet18. It consists of a series of layers connected in a specific order, used for feature extraction and classification of input images. Convolutional layers, pooling layers, fully connected layers, and the output layer are the core components of ResNet18 arranged sequentially: convolutional layers extract local spatiotemporal features; pooling layers reduce the spatial dimensionality of feature maps and enhance translation invariance; fully connected layers map high-dimensional features to the class space; and the output layer generates the final class prediction result. The Convolutional Block Attention Module (CBAM) is a lightweight attention mechanism that adaptively weights the input feature map in terms of channel dimension and spatial location by sequentially applying channel attention and spatial attention, highlighting key information and suppressing redundant responses. The feature distribution normalization module is a component inserted into the network to normalize feature vectors, typically using batch normalization to ensure the feature distribution has zero mean and unit variance, thereby improving training stability and convergence speed. The L2 regularization module is an operation unit in the network that performs L2 norm normalization on the feature vectors, scaling the feature vector of each sample to a unit hypersphere to constrain the feature scale, enhance intra-class compactness, and facilitate subsequent distance- or probability-based modeling. The initial residual network refers to a new network structure formed by embedding a convolutional block attention module after the pooling layer and adding a feature distribution normalization module and an L2 regularization module before the fully connected layer, based on ResNet18. The loss function has not yet been replaced. The cross-entropy loss function is a standard classification loss function used to measure the difference between the model's predicted class probability distribution and the true label; a smaller value indicates a more accurate prediction, as shown below:

[0086]

[0087] in, This represents the cross-entropy loss value; This indicates the total number of drone categories; This represents the total number of current input samples; Indicates the first Index of each sample; The unnormalized class score (logits) of the i-th sample output by the target residual network is a vector of length equal to the number of classes, where each element corresponds to the original confidence of a class.

[0088] The contrast center loss function is an auxiliary loss function that calculates the Euclidean distance between each sample feature and its class center, simultaneously bringing similar samples closer together and pushing away dissimilar samples, thereby achieving the effect of intra-class aggregation and inter-class separation in the feature space. It is expressed as follows:

[0089]

[0090] in, Indicates the center of comparison loss value; This indicates the number of samples acquired in a single training session; This represents the UAV feature vector obtained by extracting the i-th sample through the target residual network; Indicates the first The feature centers of each category (i.e., the category center vector of the category to which the i-th sample belongs, where...) (This is the true category label of the sample). This represents the regularization weight coefficient, used to balance the contribution ratio of intra-class compact terms and inter-class separation terms.

[0091] The improved loss function is a new loss function formed by linearly combining the cross-entropy loss function and the contrast center loss function in a weighted manner. It takes into account both classification accuracy and feature discriminativeness, and is expressed as follows:

[0092]

[0093] Where L refers to the improved loss function.

[0094] The loss calculation function is the function used to calculate the total loss of the current batch of samples during model training. The target residual network uses the above-mentioned improved loss function as its loss calculation function to guide parameter updates.

[0095] Please refer to Figure 2 , Figure 2This is a schematic diagram of the target residual network provided in Embodiment 1 of the UAV recognition method based on category-specific modeling in this application. From top to bottom, the structure is as follows: The input layer receives a time-frequency image of size 1×224×224 as input; then it passes through a 7×7 convolutional layer (stride 2, output channels 64), followed by batch normalization and ReLU activation function; then it passes through a 3×3 max pooling layer (stride 2) for spatial downsampling; next are four residual block groups, each containing multiple 3×3 convolutional layers, corresponding to 64, 128, 256 and 512 channels respectively, and repeated four times within each group for progressively extracting high-level features; after the last residual block, a convolutional block attention module is connected, which is divided into two sub-modules: the left side is the channel attention part, which is composed of global leveling... After average pooling and global max pooling, the concatenation is passed through a multilayer perceptron (compression ratio of 16) and a sigmoid activation function to generate channel weights, which weight the feature map by channel dimension. The right side is the spatial attention part, which concatenates the average pooling and max pooling results and inputs them into a 7×7 convolutional layer, then passes through a sigmoid activation function to generate a spatial mask, which weights the feature map by spatial dimension. The two attention outputs are multiplied by the original feature map to obtain the enhanced feature representation. Next is a global average pooling layer, which compresses the feature map into a 512-dimensional vector. Then it goes through a flattening layer, batch normalization (1d, 512-dimensional), L2 normalization, and a dropout layer (dropout probability of 0.3) to prevent overfitting. Finally, it is connected to the output layer through a fully connected layer to complete the classification task.

[0096] This embodiment integrates the CBAM attention mechanism into layer 4 of the ResNet18 backbone network (in ResNet18, the network mainly consists of five parts: an initial convolutional layer, four subsequent groups of convolutional layers, referred to as layers 1 to 4, and a final global average pooling layer and a fully connected output layer). In the feature extraction process, after ResNet18's layer 4 outputs a 512-dimensional feature map, it first undergoes feature enhancement via the CBAM module. Channel attention learns important frequency components in the spectrogram, while spatial attention identifies key time-frequency regions. Subsequently, batch normalization, L2 normalization, and Dropout operations further optimize feature quality. Applying this design strategy, on the one hand, higher-level features are more suitable for attention mechanism learning, effectively avoiding computational redundancy from repeatedly applying attention to lower-level features; on the other hand, features enhanced by CBAM improve the optimization efficiency of subsequent contrastive loss and the discriminative power of Gaussian mixture model open-set recognition features, synergistically optimizing the classification accuracy of known categories and the rejection capability of unknown categories, thereby comprehensively improving the open-set recognition performance of UAVs.

[0097] This embodiment embeds a convolutional block attention module after the pooling layer, enabling the network to adaptively focus on the most discriminative time-frequency regions and channels in the time-frequency map, thus suppressing background noise interference. Feature distribution normalization and L2 regularization are introduced before the fully connected layer, making the extracted UAV feature vectors more compact and consistent in scale, facilitating subsequent modeling. Simultaneously, the cross-entropy loss and contrast center loss are combined into an improved loss function, which not only optimizes classification accuracy but also explicitly constrains the clustering of similar samples and the separation of dissimilar samples in the feature space. This improves the model's stability in recognizing known categories and its ability to reject unknown categories, thereby enhancing the overall open-set recognition performance.

[0098] As an example, the training steps of the target residual network include: acquiring a training dataset and a validation dataset, both of which include radio frequency (RF) signal samples and time-frequency map samples converted from the RF signal samples, wherein the RF signal samples contain known class labels; dividing the time-frequency map samples into multiple sample batches according to a preset batch size; using the target residual network to be trained as the initial network, sequentially inputting each sample batch into the initial network to obtain a prediction result; calculating the loss value between the prediction result and the corresponding known class label; adjusting the parameters of the target residual network to be trained using a backpropagation algorithm based on the loss value, and returning to the step of sequentially inputting each sample batch into the target residual network to be trained to obtain a prediction result, until the number of iterations reaches a preset number of training rounds to obtain a reference network; calculating the recognition accuracy of the reference network based on the validation dataset; when the recognition accuracy is less than a preset accuracy threshold, returning to the step of sequentially inputting each sample batch into the target residual network to be trained to obtain a prediction result; when the recognition accuracy is greater than or equal to the preset accuracy threshold, using the reference network as the trained target residual network.

[0099] The training dataset refers to a set of labeled data used to train the target residual network. It includes radio frequency (RF) signal samples collected by several known UAV models within the monitoring area, along with their corresponding time-frequency map samples generated by short-time Fourier transform. Each sample is associated with a specific category label. The validation dataset is a separate set of data, independent of the training dataset, used to evaluate the model's generalization performance. It is structured similarly to the training dataset, including RF signal samples and corresponding time-frequency map samples with known class labels, but it does not participate in network parameter updates. RF signal samples are raw I / Q (In-phase and Quadrature) sampling data intercepted from the communication links of actual flying UAVs, representing uplink non-payload control signals or downlink image transmission signals transmitted by a specific individual within a specific time period. Time-frequency map samples are two-dimensional energy distribution images generated by applying a short-time Fourier transform to each RF signal sample, serving as input data for the target residual network. Known class labels are manually assigned category identifiers for each RF signal sample or time-frequency map sample, indicating the UAV model to which the sample belongs, covering only known categories predefined during the training phase. The preset batch size refers to the number of samples input to the network each time during training. This value is set before training to balance memory usage and gradient estimation stability. A sample batch refers to several subsets obtained by dividing all time-frequency map samples in the training dataset according to the preset batch size. Each subset contains a fixed number of samples for one forward and backward propagation calculation. The prediction result refers to the predicted class score of the target residual network outputting the class probability distribution or feature vector of each time-frequency map sample in the current sample batch after mapping by the classifier head. The loss value is a scalar value calculated using an improved loss function (a weighted sum of cross-entropy loss and contrast center loss) based on the prediction result and the corresponding known class label, used to measure the degree of error in the current network's prediction. The number of iterations refers to the number of sample batch processing rounds completed since the start of training; each sample batch processed is counted as one iteration. The preset number of training rounds refers to the maximum number of times the complete training dataset can be traversed during the entire training process. This value is set before training starts and is used as one of the termination conditions of the main loop. The reference network refers to the current version of the target residual network obtained after completing a preset number of training epochs. Its parameters have been updated through multiple rounds of backpropagation and are used for subsequent performance evaluation on the validation dataset. Recognition accuracy refers to the proportion of samples on the validation dataset where the reference network correctly predicts known class labels out of the total number of samples in the validation set; it is used to measure the model's classification performance. The preset accuracy threshold is a performance standard set before training. If the reference network's recognition accuracy on the validation dataset reaches or exceeds this value, training is considered successful and stops; otherwise, iterative optimization continues.

[0100] First, training and validation datasets are divided from the collected UAV communication data. Each dataset contains radio frequency signal samples with known class labels and their corresponding time-frequency map samples, and they are kept separate to avoid evaluation bias. Second, all time-frequency map samples in the training dataset are divided into multiple sample batches according to a preset batch size to efficiently complete gradient calculations with limited GPU memory. Then, after initializing the target residual network to be trained, each sample batch is sequentially fed into the network for forward propagation to obtain prediction results. Finally, based on these prediction results and the true known class labels, an improved loss function is used to refine the calculations. The loss function is used to calculate the loss value, and the network parameters are updated using the backpropagation algorithm. This process is repeated until the number of times the complete training set is traversed reaches the preset number of training epochs, thus obtaining a reference network. Next, the reference network is input with a validation dataset, and the proportion of samples correctly predicted is counted to obtain the recognition accuracy. If the recognition accuracy does not reach the preset accuracy threshold, it indicates that the model has not yet converged sufficiently or its generalization ability is insufficient, so a new round of training iterations continues. If the recognition accuracy has reached or exceeded the threshold, it indicates that the model performance meets the requirements, training stops, and the current reference network is finally used as the target residual network after training. This approach ensures the model's classification accuracy on known categories and prevents overfitting through an independent validation mechanism, ensuring its reliability in open environments.

[0101] Step S30: Invoke the corresponding Gaussian mixture model constructed for each known class of UAV during the training phase, and calculate the log-likelihood value of the UAV feature vector under each Gaussian mixture model.

[0102] It's important to note that the known drone classes refer to several drone model categories predefined and labeled during the training phase. Each category represents a specific model, not a single individual drone. The system constructs a separate discrimination model for each category. A Gaussian Mixture Model (GMM) is a probabilistic generative model composed of a weighted combination of multiple Gaussian distributions. It's used to fit the complex, multimodal distribution of the drone feature vector corresponding to a known drone class in the feature space. Its parameters are learned from the training samples of that category using the Expectation-Maximization (EM) algorithm. The log-likelihood value is the natural logarithm of the probability density value generated by a Gaussian Mixture Model after substituting the drone's feature vector into it. It measures the degree of matching between the feature vector and the corresponding known drone class; a higher value indicates a greater likelihood of belonging to that class.

[0103] Understandably, the system calls Gaussian mixture models built for each known class of drone during the training phase. The drone feature vector output by the target residual network of the current sample to be identified is sequentially input into each Gaussian mixture model. The probability density value of the drone feature vector generated by each model is calculated, and its natural logarithm is taken to obtain the log-likelihood value of the drone feature vector under each Gaussian mixture model.

[0104] The Gaussian mixture model is constructed independently for each known class of UAVs during the training phase: First, all training samples under that class are processed through a target residual network to extract the corresponding UAV feature vectors; then, using these feature vectors as observation data, the EM algorithm is used to iteratively optimize the parameters of the Gaussian mixture model, including the mean, covariance matrix, and mixing coefficients of each Gaussian component, until the log-likelihood converges; finally, a probabilistic generative model composed of a weighted combination of multiple Gaussian distributions is obtained, which is used to characterize the multimodal and non-uniform distribution characteristics of the features of this class of UAVs in the feature space.

[0105] Step S40: Based on the maximum value among the log-likelihood values ​​and the preset adaptive threshold, a preliminary discrimination result is obtained. The preset adaptive threshold is determined during the training phase.

[0106] It should be noted that the preset adaptive threshold refers to a dynamic discrimination boundary value determined through statistical analysis (e.g., selecting low quantile intervals, grid search, or combining validation set performance) during the training phase, based on the log-likelihood distribution calculated using a Gaussian mixture model of all known UAV classes for training or validation samples. This threshold is used to distinguish between known and unknown classes, and it can be automatically adjusted according to the data distribution characteristics rather than remaining fixed. The preliminary discrimination result refers to the initial classification decision obtained by comparing the maximum value among the log-likelihood values ​​with the preset adaptive threshold: if the maximum value is greater than or equal to the threshold, the sample is determined to belong to the known UAV class corresponding to the maximum log-likelihood value; if the maximum value is less than the threshold, the sample is determined not to belong to any known class.

[0107] As an example, the step of obtaining a preliminary discrimination result based on the maximum value in the log-likelihood value and a preset adaptive threshold includes: when the maximum value in the log-likelihood value is greater than or equal to the preset adaptive threshold, a preliminary discrimination result is obtained that the drone is a known class drone, and the known class drone belongs to the known drone class corresponding to the Gaussian mixture model to which the maximum value belongs; when the maximum value is less than the preset adaptive threshold, a preliminary discrimination result is obtained that the drone is an unknown class drone.

[0108] Known drones refer to drones whose models have appeared in the training phase and for which corresponding Gaussian mixture models have been built. The system can identify them as a specific known model through model matching. Unknown drones refer to drones whose models have not appeared in the training phase and for which no corresponding Gaussian mixture models have been built. Their radio frequency signal characteristics differ significantly from the distribution of all known classes, and therefore they are determined by the system not to belong to any known category.

[0109] The log-likelihood value reflects the probability that the UAV feature vector is generated by a Gaussian mixture model of a known class. The higher the value, the more closely the sample matches the feature distribution of the corresponding class. The preset adaptive threshold is a reasonable lower limit set during the training phase based on the log-likelihood distribution of known class samples. It represents the minimum confidence level of "acceptable as a known class". Therefore, when the maximum log-likelihood value reaches or exceeds the threshold, it indicates that the sample is sufficiently similar to a known class and can be judged as a UAV of that known class. Conversely, if the maximum value is lower than the threshold, it indicates that the sample deviates significantly from the distribution of all known classes and is very likely from a model not seen during training. Therefore, it is judged as an unknown class UAV.

[0110] The steps for determining the preset adaptive threshold include: during the training phase, acquiring training samples of all known UAV types; calculating the log-likelihood value based on the training samples and the Gaussian mixture model to form a maximum log-likelihood evidence set; performing empirical distribution estimation on the maximum log-likelihood evidence set to construct a conservative candidate domain; performing uniform grid sampling at preset intervals within the conservative candidate domain to generate multiple discrete candidate thresholds; obtaining a verification feature vector and its corresponding log-likelihood value set based on the verification dataset and the target residual network; using each candidate threshold as a criterion, comparing the log-likelihood value of the verification feature vector with the candidate threshold, evaluating the open set recognition accuracy corresponding to each candidate threshold, and selecting the candidate threshold with the highest accuracy as the initial threshold; calculating the first mean of the maximum log-likelihood evidence set and the second mean of the log-likelihood value set of the verification feature vector; when the second mean is less than the first mean, adjusting the initial threshold to obtain a preset adaptive threshold; when the second mean is greater than or equal to the first mean, using the initial threshold as the preset adaptive threshold.

[0111] Training samples refer to the time-frequency map samples of UAV radio frequency signals with known class labels used to construct Gaussian mixture models during the training phase. These samples are then used to extract UAV feature vectors through a target residual network, with each sample belonging to a known UAV class. The maximum log-likelihood evidence set refers to the set of maximum log-likelihood values ​​for each training sample after calculating its log-likelihood value under the Gaussian mixture model for each known UAV class. This set characterizes the typical response strength of known class samples under the model. The conservative candidate region refers to a numerical range defined within the low quantile interval (e.g., 1% to 10% quantile) of the empirical distribution of the maximum log-likelihood evidence set. This region focuses on the tails of the distribution and low-confidence areas easily confused with unknown classes, used to prioritize the search for robust thresholds. The preset interval refers to the fixed step size set during uniform sampling within the conservative candidate region, used to control the density and granularity of candidate thresholds. The candidate threshold refers to a series of discrete values ​​generated at preset intervals within the conservative candidate region. Each value serves as a possible discrimination boundary, used to subsequently evaluate open-set recognition performance on the validation set. The validation dataset refers to a set of test data independent of the training samples, containing known-class drones and background or unknown-class samples. It is used to evaluate the model's generalization ability and open-set discrimination performance under different candidate thresholds. The validation feature vector is the drone feature vector output after inputting time-frequency map samples from the validation dataset into the trained target residual network. It is used to calculate the log-likelihood value and perform threshold validation. The log-likelihood set is the set of all log-likelihood values ​​obtained by substituting each validation feature vector into a Gaussian mixture model of all known drone classes. For each validation sample, the maximum value is usually extracted for discrimination. Open-set recognition accuracy refers to the overall proportion of known-class samples and unknown-class samples correctly identified by the system on the validation dataset, comprehensively reflecting the model's classification and rejection capabilities in open environments. The initial threshold is the candidate threshold that maximizes the open-set recognition accuracy on the validation dataset among all candidate thresholds, serving as the benchmark for threshold selection. The first mean is the arithmetic mean of all elements in the maximum log-likelihood evidence set, reflecting the average confidence level of known-class samples during the training phase. The second mean refers to the arithmetic mean of the set of maximum log-likelihood values ​​corresponding to each of the validation feature vectors, reflecting the average response strength of the model to known class samples during the validation phase.

[0112] First, during the training phase, training samples of all known UAV classes are collected, and each sample is input into the corresponding Gaussian mixture model. The log-likelihood value of each sample under all models is calculated, and the maximum value is retained for each sample to form the maximum log-likelihood evidence set. Second, empirical distribution statistics are performed on all values ​​in this set to identify their low quantile regions, and a conservative candidate domain is constructed based on this. The purpose is to focus on low-confidence regions that are easily confused with unknown classes to improve the rejection robustness of the threshold. Then, uniform grid sampling is performed at preset intervals within the conservative candidate domain to generate a set of discrete candidate thresholds. Next, the validation feature vector is extracted through the target residual network using the validation dataset, and the log-likelihood value of each validation feature vector under each Gaussian mixture model is calculated. The maximum value is taken to obtain the corresponding log-likelihood value set. Subsequently, each candidate threshold is used as a discrimination boundary. The maximum log-likelihood value of the validation feature vector is compared with the threshold. The proportion of correctly identifying known classes and correctly rejecting unknown classes on the validation set is statistically analyzed, i.e., the open-set recognition accuracy. The candidate threshold with the highest accuracy is selected as the initial threshold. Next, the average of the maximum log-likelihood evidence set is calculated to obtain the first mean, and the average of the maximum log-likelihood values ​​of the validation feature vector set is calculated to obtain the second mean. If the second mean is less than the first mean, it indicates a decrease in the overall confidence of the validation data, possibly indicating a domain shift. In this case, the initial threshold is adjusted downwards by a certain margin to avoid over-rejection of known classes, thus obtaining the final preset adaptive threshold. If the second mean is greater than or equal to the first mean, the data distribution is considered stable, and the initial threshold is directly used as the preset adaptive threshold. This approach allows for the selection of the optimal threshold based on validation performance and dynamic correction based on the differences between training and validation distributions, improving the model's adaptability in open environments.

[0113] This embodiment introduces a GMM-based adaptive threshold into the open set recognition framework. For any test feature... First, the class-conditional log-likelihood is calculated for each category using the GMM, and the evidence with the highest value is taken as the sufficient statistic for open-set discrimination; when this evidence exceeds a threshold... If the evidence is within the known class, it is classified as such; otherwise, it is marked as an unknown class. Simultaneously, the difference between the evidence and the threshold is mapped to a continuous confidence level within the interval [0, 1] using a sigmoid function to obtain an interpretable probabilistic output. The calculation formula is as follows:

[0114]

[0115]

[0116]

[0117] in, Indicates sample The maximum log-likelihood evidence score It iterates through the indices of all known categories. It is an eigenvector In the The probability density generated under the GMM for each category. To preset the adaptive threshold, For the sample The final predicted label (classification decision result), when The output will be a known class if the condition is met, otherwise it will be an unknown class. The continuous confidence score, derived from evidence and a threshold, is used to quantify the model's confidence in distinguishing known classes. This is the Sigmoid function.

[0118] The threshold determination process comprises two stages. Stage one involves estimating the empirical distribution of the maximum log-likelihood evidence in the training set, constructing a conservative candidate domain from the very low to the low quantile intervals, focusing the search on low-density, easily confused regions at the tails of the distribution. Uniform grid sampling is then performed within this interval to form a discrete candidate set, and an exhaustive evaluation is conducted with the validation set accuracy as the objective function to obtain the optimal operating threshold. Stage two, to enhance robustness under domain shift conditions, further compares the mean differences between the training and testing evidence distributions: when a significant negative shift is detected, a conservative correction strategy is employed to lower the threshold; otherwise, it reverts to the conventional setting based on standard quantiles.

[0119] Step S50: Call the feature matrix of the known class drone samples obtained in the training phase and the preset retrieval threshold, and calculate the minimum feature distance between the drone feature vector corresponding to the unknown class drone in the preliminary discrimination result and all samples in the feature matrix.

[0120] It should be noted that "known class drone samples" refers to the time-frequency map samples corresponding to drone radio frequency signals with known class labels used to build the model during the training phase. These samples are then used to extract drone feature vectors through the target residual network, with each sample representing a typical feature instance of a known model. The feature matrix is ​​a matrix composed of rows of drone feature vectors from all known class drone samples, with each row corresponding to a feature vector of a training sample, used for subsequent distance calculation and nearest neighbor matching. The preset retrieval threshold is an upper limit determined during the training phase by statistically analyzing the minimum Euclidean distance between every pair of known class drone samples and taking its 95th quantile. This threshold is used to determine whether a sample initially classified as unknown is close enough to be classified as known, thus deciding whether to "retrieve" it as known. The minimum feature distance is the minimum value selected from the Euclidean distances calculated between the feature vector of a drone initially classified as unknown and the feature vectors of every known class drone sample in the feature matrix. This minimum value is used to measure the nearest neighbor similarity between the unknown sample and the overall known class.

[0121] Understandably, firstly, the system calls the feature matrix of known class drone samples generated during the training phase (this matrix is ​​formed by stacking drone feature vectors extracted from all training samples by the target residual network row by row) and the preset retrieval threshold; secondly, for each drone feature vector that is judged as an unknown class drone in the preliminary discrimination results, the system iterates through the Euclidean distance between the feature vector and each row in the feature matrix (i.e., each known class drone sample), calculates and records all distance values ​​one by one; then, it finds the smallest one among these distance values ​​as the minimum feature distance between the unknown discrimination sample and the set of known class samples.

[0122] As an example, the steps for obtaining the feature matrix and the preset recovery threshold include: during the training phase, acquiring training samples of all known UAV classes; extracting features from each training sample using the target residual network to obtain a sample feature vector corresponding to each training sample; classifying and organizing the sample feature vectors according to the known class labels of the training samples, and combining all sample feature vectors corresponding to the same known class label to obtain a feature matrix of the known UAV samples; calculating the pairwise Euclidean distance between any two sample feature vectors in the feature matrix to obtain a distance matrix, wherein the diagonal elements of the distance matrix are set to infinity; selecting the minimum distance between each sample feature vector and all other sample feature vectors from the distance matrix to obtain a minimum distance set; and using the 95th quantile of the minimum distance set as the preset recovery threshold.

[0123] A sample feature vector is a fixed-dimensional real-valued vector output by the target residual network after inputting training samples of each known class of UAV during the training phase. It represents the position of the sample in the feature space. Pairwise Euclidean distance is the linear distance between any two distinct sample feature vectors in the feature matrix, calculated as the square root of the sum of the squares of the differences in each dimension. It measures the similarity between the two samples in the feature space. The diagonal elements in the distance matrix, where the row and column indices are the same, correspond to the distance between the same sample feature vector and itself. These are explicitly set to infinity in the calculation to prevent incorrectly identifying itself as the nearest sample during subsequent nearest neighbor search. The minimum distance is the smallest value among the pairwise Euclidean distances of each sample feature vector in the feature matrix with all other sample feature vectors (excluding itself). It represents the distance of that sample to the next nearest neighbor in the known class set. The minimum distance set refers to the set of minimum distances corresponding to each sample feature vector in the feature matrix. It is used to statistically analyze the closeness between samples within a known class and serves as the basis for setting the retrieval threshold.

[0124] Classification and organization process: Group all sample feature vectors by category, and concatenate all sample feature vectors belonging to the same known class label by row to form the feature matrix of that category. The feature matrices of all categories together constitute the feature matrix of the known class of UAV samples.

[0125] The nearest neighbor retrieval strategy proposed in this embodiment is a post-processing mechanism based on feature space distance. After using the Gaussian Mixture Model (GMM) to perform open-set discrimination on the performance representation in the open-set recognition task, some known-class test samples may be classified as unknown-class due to their low log-likelihood values. To further improve classification reliability, a nearest neighbor retrieval mechanism is introduced to perform secondary discrimination on these potentially known-class samples. First, features are extracted from the known-class samples in the training set, a feature matrix is ​​constructed, and the pairwise Euclidean distances between samples are calculated. By setting the diagonal elements of the distance matrix to infinity to avoid self-distance interference, the minimum distance from each training sample to all other samples is calculated, and the 95th percentile is used as the retrieval threshold. When the minimum feature distance between a sample in the test set classified as an unknown-class and a known-class sample is less than the retrieval threshold, it is reclassified as the corresponding known-class. This mechanism can effectively identify boundary samples misclassified by the GMM, especially those samples whose feature representations are close to the known class but are incorrectly classified as unknown-class, reducing the misclassification rate of samples.

[0126] Step S60: When the minimum feature distance is less than the preset recovery threshold, the unknown type of UAV is reclassified as the known type of UAV corresponding to the minimum feature distance.

[0127] Understandably, when the minimum feature distance is less than the preset recovery threshold, the system reclassifies the sample initially identified as an unknown drone as a known drone. Specifically, the minimum feature distance is the minimum value obtained after calculating the Euclidean distance between the current drone's feature vector and the feature vectors of all known drone samples in the feature matrix. The system records the index position corresponding to this minimum value during the distance calculation process to accurately locate the sample closest to it in the feature matrix, i.e., the nearest neighbor sample (referring to the known class training sample with the smallest Euclidean distance to the current sample in the feature space). Subsequently, the system directly uses the known class label associated with this nearest neighbor sample during the training phase as the final category of the current sample, thereby completing the recovery of the sample originally misclassified as an unknown drone.

[0128] This embodiment provides a UAV identification method based on category-specific modeling. First, the radio frequency signals of UAVs within the monitoring area are acquired and subjected to a short-time Fourier transform to generate a time-frequency map that simultaneously characterizes the signal's temporal evolution and frequency components, providing a two-dimensional input rich in discriminative information for subsequent identification. Next, this time-frequency map is input into a target residual network for feature extraction, yielding a compact and discriminative UAV feature vector. This network is an improvement on ResNet18, maintaining deep feature extraction capabilities while enhancing attention to key time-frequency regions and optimizing the feature space structure. Subsequently, a Gaussian mixture model, constructed separately for each known UAV class during the training phase, is invoked to calculate the UAV's... The system calculates the log-likelihood values ​​of feature vectors under each model and obtains preliminary discrimination results by comparing the maximum log-likelihood value with a preset adaptive threshold. The adaptive threshold is dynamically determined based on the likelihood distribution of training and validation data, improving the stability of open-set discrimination. To further reduce the misclassification of known boundary samples as unknown classes, the system calls upon the feature matrices of known class drone samples constructed during the training phase and a preset retrieval threshold to calculate the minimum feature distance between the feature vectors of drones initially classified as unknown and all known class samples. If this distance is less than the preset retrieval threshold, the drone is reclassified as belonging to the known class of its nearest neighbor, effectively correcting the conservative rejection that might occur with the Gaussian mixture model. This embodiment, through a multi-stage collaborative discrimination mechanism, can accurately identify known class drones and effectively distinguish unknown class drones in open environments.

[0129] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 , Figure 3This is a flowchart illustrating the second embodiment of the UAV recognition method based on category-specific modeling of this application. The target residual network includes convolutional layers, pooling layers, convolutional block attention modules, feature distribution normalization modules, L2 regularization modules, fully connected layers, and an output layer. Step S20 of the UAV recognition method based on category-specific modeling includes steps S21 to S26:

[0130] Step S21: Input the time-frequency map into the target residual network, and extract features from the time-frequency map through the convolutional layer to obtain local features.

[0131] It should be noted that local features refer to the low- or mid-level feature maps output by the convolutional layers in the target residual network after performing sliding window convolution operations on the input time-frequency map. These features retain fine-grained structural information such as texture, edges, and energy changes in the local time-frequency region of the time-frequency map, have spatial location correspondence, and have not yet undergone global pooling or high-level semantic abstraction.

[0132] Step S22: The local features are downsampled through the pooling layer to obtain preliminary features.

[0133] It should be noted that preliminary features refer to the feature maps obtained after the pooling layer performs downsampling operations on local features. Their spatial dimensions are reduced compared to local features, preserving the main discriminative responses and suppressing redundant information. They are intermediate feature representations after preliminary scale compression and invariance enhancement.

[0134] Step S23: Enhance the preliminary features by performing channel dimension and spatial dimension feature enhancement on the convolutional block attention module to obtain enhanced features.

[0135] It's important to note that the channel dimension refers to the depth direction of the feature map in a convolutional neural network. Each channel in the feature map corresponds to a specific type of feature response (such as a frequency pattern or modulation feature). Operations in the channel dimension aim to adjust the importance weights of each channel. The spatial dimension refers to the height and width of the feature map, corresponding to the time-frequency coordinates in the time-frequency map. Operations in the spatial dimension are used to identify which specific time-frequency regions are more discriminative for the current recognition task. Enhanced features refer to the new feature map output after the initial features are processed sequentially by the channel attention submodule and the spatial attention submodule of the convolutional block's attention module. Channels with strong discriminative power are given higher weights, key time-frequency regions are highlighted, and redundant or noisy responses are suppressed, thus achieving targeted enhancement of feature representations in both the channel and spatial dimensions.

[0136] Understandably, please refer to Figure 4 , Figure 4This is a schematic diagram of the processing flow of the convolutional block attention module in the UAV recognition method based on category-specific modeling provided in Embodiment 2 of this application. The input is a feature map (preliminary feature) of size C×H×W, where C represents the number of channels, and H and W represent the height and width, respectively. This input feature map is simultaneously fed into the channel attention module on the left and the spatial attention module on the right. In the channel attention module, the input feature map first undergoes global average pooling and global max pooling operations to compress the spatial dimension into two C×1×1 vectors. These two vectors are then input into a shared convolutional layer (containing a convolutional 2d + ReLU + convolutional 2d structure) for nonlinear transformation. The output is then element-wise compared with the two pooling results. The input feature map is then processed by a Sigmoid activation function to generate a C×1×1 channel weight matrix. In the spatial attention module, the input feature map is first subjected to average pooling and max pooling along the channel dimension to obtain two 1×H×W feature maps. These two maps are then concatenated along the channel dimension to form a 2×H×W feature map, which is then fused into a 7×7 convolutional layer. Finally, a 1×H×W spatial weight matrix is ​​generated by the Sigmoid activation function. Subsequently, the channel attention weight matrix and the spatial attention weight matrix are multiplied element-wise with the original input feature map to achieve dual weighted enhancement of the channel and spatial dimensions. The final output is an enhanced feature map with the same size of C×H×W. This approach allows the network to adaptively focus on the most discriminative feature channels and spatiotemporal locations, improving the robustness of UAV signal recognition under weak or interfered conditions in complex electromagnetic environments.

[0137] Step S24: The enhanced features are processed by the feature distribution standardization module to obtain standardized features.

[0138] It should be noted that standardized features refer to enhanced features that have been processed by a feature distribution standardization module (such as batch normalization) so that each dimension has zero mean, unit variance, or uniform scale.

[0139] Understandably, the process involves several steps. First, the enhanced features are fed into a feature distribution standardization module. For each channel, the mean and standard deviation of the channel's feature values ​​are calculated across all samples in the current batch. Second, the original feature values ​​are subtracted from the mean of the corresponding channel and divided by the standard deviation of that channel to achieve zero mean and unit variance normalization. Then, the normalized result is multiplied by a learnable scaling factor and a learnable offset is added to preserve the network's flexibility in representing the feature distribution. Finally, the standardized features are output after the above processing. This approach aims to mitigate feature distribution shift issues caused by attention mechanisms or deep layer stacking, accelerate model convergence, and improve the stability and effectiveness of subsequent L2 regularization and Gaussian mixture modeling.

[0140] Step S25: The standardized features are regularized by the L2 regularization processing module to obtain regularized features.

[0141] It should be noted that regularized features refer to the feature vectors obtained after the standardized features have been processed by the L2 regularization module. The feature vector of each sample is divided by its own L2 norm (i.e., Euclidean length) so that the vector lies on a unit hypersphere in the feature space and has a unit length.

[0142] Understandably, the process involves several steps. First, the standardized features are fed into the L2 regularization module, where the L2 norm of each sample's feature vector is calculated by summing the squares of all dimensions of the vector and then taking the square root. Second, each element of the feature vector is divided by the calculated L2 norm to normalize the vector length. Finally, unit-length vectors corresponding to all samples are output, forming the regularized features. This is done to constrain the features to a unit hypersphere, eliminating the impact of scale differences on subsequent Gaussian mixture modeling and distance metrics. This results in a more compact intra-class distribution and more reliable inter-class discrimination, thereby improving the stability and accuracy of open set recognition.

[0143] Step S26: The regularized features are dimensionally mapped and integrated through the fully connected layer to obtain the UAV feature vector, and the UAV feature vector is output through the output layer.

[0144] Understandably, firstly, matrix multiplication is performed between the pre-defined weight matrix and regularized features in the fully connected layer to map the originally high-dimensional regularized features to a pre-defined dimension. This is done to eliminate redundant information in the high-dimensional features and make the feature dimension adaptable to the processing requirements of the subsequent recognition model. Then, the fully connected layer integrates the mapped features. Specifically, through non-linear transformations of neurons within the layer, the scattered key feature information related to the drone category in the mapped features is fused and summarized to form a feature set with global discriminative power. This is because the previous regularized features were still local discrete features; integration strengthens the feature's ability to represent the drone category. Finally, the drone feature vector formed after dimension mapping and feature integration is transmitted to the output layer, which outputs the drone feature vector.

[0145] This embodiment first inputs the time-frequency image into the target residual network, extracts local features through convolutional layers, and captures fine-grained structural information of local time-frequency regions in the time-frequency image. Then, pooling layers downsample the local features to obtain preliminary features, which reduce redundancy and enhance robustness to small shifts while retaining the main responses. Next, the convolutional block attention module sequentially adjusts the channel and spatial dimensions of the preliminary features to strengthen the discriminative feature channels and focus on key time-frequency regions, thereby obtaining enhanced features that effectively suppress noise interference. Then, the feature distribution normalization module normalizes the enhanced features to obtain standardized features, making the feature distribution more stable and beneficial for subsequent processing. Then, the L2 regularization module constrains the standardized features with the unit norm to obtain regularized features, eliminating scale differences and facilitating unified modeling. Finally, the fully connected layer linearly maps and integrates the regularized features to output a fixed-dimensional UAV feature vector, providing a structurally consistent and highly discriminative feature representation for subsequent category-specific modeling.

[0146] This application also provides a drone identification device based on category-specific modeling, please refer to... Figure 5 The category-specific modeling-based drone identification device includes:

[0147] The time-frequency diagram generation module 10 is used to acquire the radio frequency signal of the UAV in the monitoring area and perform a short-time Fourier transform on the radio frequency signal to obtain a time-frequency diagram.

[0148] Feature extraction module 20 is used to input the time-frequency map into the target residual network for feature extraction to obtain the UAV feature vector. The target residual network is an improvement on ResNet18.

[0149] The likelihood calculation module 30 is used to call the corresponding Gaussian mixture model constructed for each known class of UAV during the training phase, and calculate the log-likelihood value of the UAV feature vector under each Gaussian mixture model.

[0150] The preliminary discrimination module 40 is used to compare the maximum value in the log-likelihood values ​​with a preset adaptive threshold to obtain a preliminary discrimination result. The preset adaptive threshold is determined during the training phase.

[0151] The distance calculation module 50 is used to call the feature matrix of the known class drone samples obtained in the training phase and the preset recovery threshold, and to calculate the minimum feature distance between the drone feature vector corresponding to the unknown class drone in the preliminary discrimination result and all samples in the feature matrix;

[0152] The secondary discrimination module 60 is used to reclassify the unknown type drone as the known type drone corresponding to the minimum feature distance when the minimum feature distance is less than the preset recovery threshold.

[0153] The category-specific modeling-based drone identification device provided in this application, employing the category-specific modeling-based drone identification method described in the above embodiments, can solve the technical problem of accurately identifying known types of drones and effectively distinguishing unknown types of drones in an open environment. Compared with the prior art, the beneficial effects of the category-specific modeling-based drone identification device provided in this application are the same as those of the category-specific modeling-based drone identification method provided in the above embodiments, and other technical features in the category-specific modeling-based drone identification device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0154] This application provides a drone identification device based on category-specific modeling. The drone identification device based on category-specific modeling includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the drone identification method based on category-specific modeling in the above embodiment 1.

[0155] The following is for reference. Figure 6 This document illustrates a structural schematic diagram of a drone identification device based on category-specific modeling suitable for implementing embodiments of this application. The drone identification device based on category-specific modeling in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 6 The drone identification device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0156] like Figure 6As shown, the category-specific modeling-based UAV identification device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in ROM (Read Only Memory) 1002 or a program loaded from storage device 1003 into RAM (Random Access Memory) 1004. RAM 1004 also stores various programs and data required for the operation of the category-specific modeling-based UAV identification device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. Input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, LCDs (Liquid Crystal Displays), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the category-specific modeling-based UAV identification device to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows category-specific modeling-based UAV identification devices with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0157] The category-specific modeling-based drone identification device provided in this application, employing the category-specific modeling-based drone identification method described in the above embodiments, can solve the technical problem of accurately identifying known types of drones and effectively distinguishing unknown types of drones in an open environment. Compared with the prior art, the beneficial effects of the category-specific modeling-based drone identification device provided in this application are the same as those of the category-specific modeling-based drone identification method provided in the above embodiments, and other technical features in this category-specific modeling-based drone identification device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0158] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the category-specific modeling-based drone identification method in the above embodiments.

[0159] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a UAV identification device based on category-specific modeling, the UAV identification device based on category-specific modeling performs the following actions: acquires radio frequency signals of UAVs within a monitoring area and performs a short-time Fourier transform on the radio frequency signals to obtain a time-frequency map; inputs the time-frequency map into a target residual network for feature extraction to obtain UAV feature vectors, wherein the target residual network is an improvement on ResNet18; calls the corresponding Gaussian mixture model constructed for each known class of UAV during the training phase and calculates the log-likelihood value of the UAV feature vector under each Gaussian mixture model; obtains a preliminary discrimination result based on the maximum value of the log-likelihood value and a preset adaptive threshold, wherein the preset adaptive threshold is determined during the training phase; calls the feature matrix of known class UAV samples obtained during the training phase and a preset recovery threshold, and calculates the minimum feature distance between the UAV feature vector corresponding to the unknown class UAV in the preliminary discrimination result and all samples in the feature matrix; when the minimum feature distance is less than the preset recovery threshold, the unknown class UAV is reclassified as a known class UAV corresponding to the minimum feature distance.

[0160] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0161] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described category-specific modeling-based UAV identification method. This solves the technical problem of accurately identifying known UAV classes and effectively distinguishing unknown UAV classes in an open environment. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the category-specific modeling-based UAV identification method provided in the above embodiments, and will not be repeated here.

[0162] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described category-specific modeling-based drone identification method.

[0163] The computer program product provided in this application can solve the technical problem of how to accurately identify known types of drones and effectively distinguish unknown types of drones in an open environment. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the drone identification method based on category-specific modeling provided in the above embodiments, and will not be repeated here.

[0164] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for identifying unmanned aerial vehicles (UAVs) based on category-specific modeling, characterized in that, The method includes: The radio frequency signals of the UAVs within the monitoring area are acquired, and the radio frequency signals are subjected to short-time Fourier transform to obtain a time-frequency diagram; The time-frequency graph is input into the target residual network for feature extraction to obtain the UAV feature vector. The target residual network is an improvement on ResNet18. The corresponding Gaussian mixture model constructed for each known class of UAV during the training phase is invoked, and the log-likelihood value of the UAV feature vector under each Gaussian mixture model is calculated. A preliminary discrimination result is obtained based on the maximum value among the log-likelihood values ​​and a preset adaptive threshold, wherein the preset adaptive threshold is determined during the training phase; The feature matrix of the known class drone samples obtained in the training phase and the preset recovery threshold are called, and the minimum feature distance between the drone feature vector corresponding to the unknown class drone in the preliminary discrimination result and all samples in the feature matrix is ​​calculated. When the minimum feature distance is less than the preset recovery threshold, the unknown type of UAV is reclassified as the known type of UAV corresponding to the minimum feature distance; The construction steps of the target residual network include: The ResNet18 is used as the basic network architecture, which includes convolutional layers, pooling layers, fully connected layers, and an output layer. A convolutional block attention module is embedded after the pooling layer, and a feature distribution normalization module and an L2 regularization processing module are embedded before the fully connected layer to obtain the initial residual network. By combining the cross-entropy loss function with the contrast center loss function, an improved loss function is obtained. The improved loss function is used as the loss calculation function of the initial residual network to obtain the target residual network; The steps for obtaining the feature matrix and the preset recycling threshold include: During the training phase, training samples of all known types of UAVs are acquired; The target residual network is used to extract features from each training sample to obtain a sample feature vector corresponding to each training sample. The sample feature vectors are classified and organized according to the known class labels of the training samples, and all the sample feature vectors corresponding to the same known class label are combined to obtain the feature matrix of the known class UAV samples; Calculate the pairwise Euclidean distance between any two sample feature vectors in the feature matrix to obtain a distance matrix, wherein the elements at the diagonal positions in the distance matrix are set to infinity; The minimum distance between each sample feature vector and all other sample feature vectors is selected from the distance matrix to obtain the minimum distance set; The 95th percentile of the minimum distance set is used as the preset recycling threshold.

2. The method as described in claim 1, characterized in that, The target residual network includes convolutional layers, pooling layers, convolutional block attention modules, feature distribution normalization modules, L2 regularization modules, fully connected layers, and an output layer. The step of inputting the time-frequency map into the target residual network for feature extraction to obtain the UAV feature vector includes: The time-frequency image is input into the target residual network, and features are extracted from the time-frequency image through the convolutional layer to obtain local features; The local features are downsampled using the pooling layer to obtain preliminary features; The initial features are enhanced by the convolutional block attention module in both channel and spatial dimensions to obtain enhanced features. The enhanced features are processed by the feature distribution standardization module to obtain standardized features. The standardized features are regularized using the L2 regularization processing module to obtain regularized features; The UAV feature vector is obtained by performing dimension mapping and feature integration on the regularized features through the fully connected layer, and then output through the output layer.

3. The method as described in claim 1, characterized in that, The training steps for the target residual network include: Obtain a training dataset and a validation dataset, both of which include radio frequency signal samples and time-frequency plot samples converted from the radio frequency signal samples, wherein the radio frequency signal samples contain known class labels; The time-frequency graph samples are divided into multiple sample batches according to a preset batch size; The target residual network to be trained is used as the initial network, and each batch of samples is sequentially input into the initial network to obtain the prediction results. Calculate the loss value between the prediction result and the corresponding known class label; Based on the loss value, the parameters of the target residual network to be trained are adjusted using the backpropagation algorithm, and the process of inputting each batch of samples into the target residual network to be trained in sequence to obtain the prediction result is repeated until the number of iterations reaches the preset number of training rounds to obtain the reference network. The recognition accuracy of the reference network is calculated based on the validation dataset; When the recognition accuracy is less than the preset accuracy threshold, return to the step of sequentially inputting each batch of samples into the target residual network to be trained to obtain the prediction result; When the recognition accuracy is greater than or equal to the preset accuracy threshold, the reference network is used as the target residual network after training.

4. The method as described in claim 1, characterized in that, The steps of acquiring the radio frequency signal of the UAV within the monitoring area and performing a short-time Fourier transform on the radio frequency signal to obtain a time-frequency diagram include: Acquire radio frequency signals of UAVs within the monitoring area, the radio frequency signals including uplink non-payload control signals and downlink image transmission signals; The radio frequency signal is segmented into segments by a preset sliding window function to obtain multiple local signals; Perform a Fourier transform on each of the local signals to obtain the frequency domain feature data corresponding to each local signal; The power spectral density of the radio frequency signal is calculated based on all the frequency domain feature data, and a time-frequency diagram is generated based on the power spectral density.

5. The method according to any one of claims 1 to 4, characterized in that, The step of obtaining the preliminary discrimination result based on the maximum value of the log-likelihood values ​​and the preset adaptive threshold includes: When the maximum value in the log-likelihood values ​​is greater than or equal to a preset adaptive threshold, a preliminary judgment result is obtained that the drone is a known class drone, and the known class drone belongs to the known drone class corresponding to the Gaussian mixture model to which the maximum value belongs; When the maximum value is less than the preset adaptive threshold, a preliminary judgment result is obtained that the drone is an unknown type of drone.

6. A drone identification device based on category-specific modeling, characterized in that, The device includes: The time-frequency diagram generation module is used to acquire the radio frequency signals of the UAV within the monitoring area and perform a short-time Fourier transform on the radio frequency signals to obtain a time-frequency diagram. The feature extraction module is used to input the time-frequency image into the target residual network for feature extraction to obtain the UAV feature vector. The target residual network is an improvement on ResNet18. The construction steps of the target residual network include: using ResNet18 as the basic network architecture, which includes convolutional layers, pooling layers, fully connected layers, and an output layer; embedding a convolutional block attention module after the pooling layer, and embedding a feature distribution normalization module and an L2 regularization processing module before the fully connected layer to obtain an initial residual network; combining the cross-entropy loss function and the contrast center loss function to obtain an improved loss function; and using the improved loss function as the loss calculation function of the initial residual network to obtain the target residual network. The likelihood calculation module is used to call the corresponding Gaussian mixture model built for each known class of UAV during the training phase, and calculate the log-likelihood value of the UAV feature vector under each Gaussian mixture model. The preliminary discrimination module is used to compare the maximum value in the log-likelihood values ​​with a preset adaptive threshold to obtain a preliminary discrimination result. The preset adaptive threshold is determined during the training phase. The distance calculation module is used to call the feature matrix of known class drone samples obtained in the training phase and the preset recovery threshold, and to calculate the minimum feature distance between the drone feature vector corresponding to the unknown class drone in the preliminary discrimination result and all samples in the feature matrix; the steps of obtaining the feature matrix and the preset recovery threshold include: in the training phase, obtaining training samples of all known class drones; extracting features from each training sample through the target residual network to obtain the sample feature vector corresponding to each training sample; classifying and organizing the sample feature vectors according to the known class label of the training samples, and combining all sample feature vectors corresponding to the same known class label to obtain the feature matrix of the known class drone samples; calculating the pairwise Euclidean distance between any two sample feature vectors in the feature matrix to obtain a distance matrix, wherein the diagonal elements in the distance matrix are set to infinity; selecting the minimum distance between each sample feature vector and all other sample feature vectors from the distance matrix to obtain the minimum distance set; and using the 95th quantile of the minimum distance set as the preset recovery threshold; The secondary discrimination module is used to reclassify the unknown type drone as the known type drone corresponding to the minimum feature distance when the minimum feature distance is less than the preset recovery threshold.

7. A drone identification device based on category-specific modeling, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the unmanned aerial vehicle identification method based on category-specific modeling as described in any one of claims 1 to 5.

8. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the UAV identification method based on category-specific modeling as described in any one of claims 1 to 5.

Images