Method for interference identification of unmanned aerial vehicle signals

Abstract

The application provides a method for identifying interference of unmanned aerial vehicle signals, and relates to the technical field of signal identification. The method comprises the following steps: obtaining an unmanned aerial vehicle signal interfered; adopting short-time Fourier transform to convert and process the unmanned aerial vehicle signal, and obtaining a two-dimensional time-frequency graph corresponding to the unmanned aerial vehicle signal; inputting the two-dimensional time-frequency graph into an interference type identification model, outputting an interference type to which the interference signal belongs through the interference type identification model; and inputting the two-dimensional time-frequency graph into an interference target detection model based on parallel fusion of a convolution operation path and a self-attention operation path, outputting a prediction box and a confidence score corresponding to the interference signal through the interference target detection model; wherein the identification result of the interference signal comprises the interference type, the prediction box and the confidence score, the prediction box is used for indicating the space-time position of the interference signal in the two-dimensional time-frequency graph, and the confidence score is used for indicating the confidence of the prediction box. The technical scheme provided by the application can effectively improve the accuracy of interference identification.

Description

Technical Field

[0001] This application relates to the field of signal recognition technology, and in particular to a method for identifying interference in UAV signals. Background Technology

[0002] The communication links upon which drone technology relies are susceptible to various unintentional or man-made interference signals, leading to serious consequences such as loss of remote control commands, interruption of image transmission, or even loss of aircraft connection. Therefore, effectively identifying interference affecting drone signals is a prerequisite for ensuring stable and reliable communication links and implementing targeted anti-interference measures.

[0003] Interference signals exhibit a trend towards diversification, transient changes, and wide coverage. Traditional identification methods based on manual feature extraction often suffer from low accuracy when facing unknown or complex modulation interference due to insufficient feature representation capabilities. While deep learning-based identification methods can improve the automation of feature extraction to some extent, their model structures are often designed for a single dimension. Furthermore, in complex interference scenarios, they only determine the presence or absence of interference signals and cannot identify the precise time point, frequency range, or other spatiotemporal parameters of the interference, resulting in low accuracy in interference identification.

[0004] Therefore, how to effectively improve the accuracy of interference identification of UAV signals is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] This application provides a method for identifying interference in UAV signals, which can effectively improve the accuracy of interference identification in UAV signals.

[0006] This application provides a method for identifying interference in UAV signals, including: Acquire jammed drone signals; The UAV signal is converted and processed using short-time Fourier transform to obtain a two-dimensional time-frequency diagram corresponding to the UAV signal; The two-dimensional time-frequency image is input into the interference category identification model, which outputs the interference type to which the interference signal belongs; and the two-dimensional time-frequency image is input into the interference target detection model based on the parallel fusion of convolution operation path and self-attention operation path, which outputs the prediction box and confidence score corresponding to the interference signal. The identification result of the interference signal includes the interference type, the prediction box, and the confidence score. The prediction box is used to indicate the spatiotemporal position of the interference signal in the two-dimensional time-frequency diagram, and the confidence score is used to indicate the confidence of the prediction box.

[0007] According to the interference identification method for UAV signals provided in this application, the step of converting the UAV signal using short-time Fourier transform to obtain a two-dimensional time-frequency diagram corresponding to the UAV signal includes: A time-localized time window function is applied to the UAV signal, and a Fourier transform is performed on the signal segment within the current time window to obtain the spectral information of the signal segment corresponding to the current time window. The current time window is slid along the time axis to obtain the spectrum diagrams of the signal segments corresponding to each of the multiple time windows; wherein, the spectrum diagrams of the signal segments corresponding to each of the multiple time windows constitute the two-dimensional time-frequency diagram, and there is overlap between adjacent time windows.

[0008] According to the interference identification method for UAV signals provided in this application, the interference target detection model includes a backbone network and a feature processing module; The step of inputting the two-dimensional time-frequency image into an interference target detection model based on the parallel fusion of convolution operation paths and self-attention operation paths, and outputting the prediction box and confidence score corresponding to the interference signal through the interference target detection model, includes: The two-dimensional time-frequency map is input into the backbone network based on the parallel fusion of convolution operation path and self-attention operation path. The backbone network performs multi-scale convolution on the two-dimensional time-frequency map and outputs the target multi-scale feature map corresponding to the two-dimensional time-frequency map. The target multi-scale features are input into the feature processing module, which performs feature interaction and fusion on the target multi-scale features and outputs the prediction box and confidence score corresponding to the interference signal.

[0009] According to the interference identification method for UAV signals provided in this application, the backbone network includes a first high-dimensional feature extraction module and a second high-dimensional feature extraction module based on the parallel fusion of convolution operation path and self-attention operation path; The two-dimensional time-frequency image is input into the backbone network based on the parallel fusion of convolutional operation paths and self-attention operation paths. The backbone network performs multi-scale convolution on the two-dimensional time-frequency image, and outputs the target multi-scale feature map corresponding to the two-dimensional time-frequency image, including: The two-dimensional time-frequency graph is input into the first high-dimensional feature extraction module, and the high-dimensional feature extraction module performs high-dimensional feature extraction on the two-dimensional time-frequency graph and outputs high-dimensional features. The high-dimensional features are input into the second high-dimensional feature extraction module, which performs parallel feature extraction based on convolution operation path and self-attention operation path, and fuses the parallel extracted features to output the target multi-scale feature map.

[0010] According to the interference identification method for UAV signals provided in this application, the second high-dimensional feature extraction module includes multiple dynamically convolutional and self-attention bottleneck modules connected in series. The high-dimensional features are input into a second high-dimensional feature extraction module that uses a parallel fusion of convolutional and self-attention operation paths. This module performs parallel feature extraction on the high-dimensional features using both convolutional and self-attention operation paths, and then fuses the extracted features to output the target multi-scale feature map, including: The high-dimensional features are input into the first dynamic convolution and self-attention bottleneck module. The high-dimensional features are extracted in parallel based on the convolution operation path and the self-attention operation path through the first dynamic convolution and self-attention bottleneck module. The extracted features are then fused to output the first multi-scale feature map. For the other dynamic convolution and self-attention bottleneck modules besides the first one, parallel feature extraction based on convolution operation path and self-attention operation path is performed on the second multi-scale feature map output by the previous dynamic convolution and self-attention bottleneck module through the other dynamic convolution and self-attention bottleneck modules, and the parallel extracted features are fused to output a third multi-scale feature map; wherein, the third multi-scale feature map output by the last dynamic convolution and self-attention bottleneck module among the other dynamic convolution and self-attention bottleneck modules is the target multi-scale feature map.

[0011] According to the interference identification method for UAV signals provided in this application, the first dynamic convolution and self-attention bottleneck module includes a convolution projection unit, a convolution unit based on the convolution operation path, and a self-attention unit based on the self-attention operation path. The high-dimensional features are input into the first dynamic convolution and self-attention bottleneck module. This module performs parallel feature extraction on the high-dimensional features based on convolutional and self-attention operation paths. The extracted features are then fused to output a first multi-scale feature map, including: The high-dimensional features are input into the convolutional projection unit, and the high-dimensional features are linearly projected through the convolutional projection unit to output the query, key, and value. The query, key, and value are respectively input into the convolutional unit and the self-attention unit. The convolutional unit performs a depthwise separable convolution operation on the query, key, and value, and outputs a convolutional feature map. The self-attention unit performs a multi-head self-attention operation on the query, key, and value, and outputs a self-attention feature map. The convolutional feature map and the self-attention feature map are fused to obtain the first multi-scale feature map.

[0012] According to the interference identification method for UAV signals provided in this application, the step of fusing the convolutional feature map and the self-attention feature map to obtain the first multi-scale feature map includes: Determine the gating coefficients corresponding to the convolutional unit and the self-attention unit respectively; Based on the gating coefficients corresponding to the convolutional unit and the self-attention unit, the convolutional feature map and the self-attention feature map are weighted and fused to obtain the first multi-scale feature map.

[0013] According to the interference identification method for UAV signals provided in this application, the feature processing module includes an attention-based intra-scale feature interaction (AIFI) module, a cross-scale feature fusion (CFFM) module, and a decoding module. The step of inputting the multi-scale features of the target into the feature processing module, performing feature interaction and fusion on the multi-scale features of the target through the feature processing module, and outputting the prediction box and confidence score corresponding to the interference signal includes: The target multi-scale feature map is input into the AIFI module, and the AIFI module performs intra-scale feature self-attention fusion on the target multi-scale feature map to output the first fused feature. Both the target multi-scale feature map and the first fused feature are input into the CFFM. The CFFM is used to fuse the target multi-scale feature map and the first fused feature to obtain the second fused feature. The second fused feature is input into the decoding module, which decodes the second fused feature based on decoupled location and content queries, and outputs the prediction box and confidence score corresponding to the interference signal.

[0014] According to the interference identification method for UAV signals provided in this application, the decoding module decodes the second fused feature based on decoupled location and content queries, and outputs a confidence score corresponding to the interference signal, including: Based on the decoupled location and content query pair, the second fused feature is decoded to output an initial confidence score; The confidence score is determined based on the initial confidence score and the IOU confidence score predicted based on IOU.

[0015] According to the interference identification method for UAV signals provided in this application, determining the confidence score based on the initial confidence score and the IOU confidence score based on IOU prediction includes: Determine the product of the initial credibility score and the IOU credibility score; The product is determined as the confidence score.

[0016] This application also provides an interference identification device for drone signals, including: Acquisition unit, used to acquire signals from interfered drones; The conversion unit is used to convert the UAV signal using short-time Fourier transform to obtain a two-dimensional time-frequency diagram corresponding to the UAV signal. The identification unit is used to input the two-dimensional time-frequency image into the interference category identification model, and output the interference type to which the interference signal belongs through the interference category identification model; and to input the two-dimensional time-frequency image into the interference target detection model based on the parallel fusion of convolution operation path and self-attention operation path, and output the prediction box and confidence score corresponding to the interference signal through the interference target detection model. The identification result of the interference signal includes the interference type, the prediction box, and the confidence score. The prediction box is used to indicate the spatiotemporal position of the interference signal in the two-dimensional time-frequency diagram, and the confidence score is used to indicate the confidence of the prediction box.

[0017] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the interference identification method for drone signals as described above.

[0018] This application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the interference identification method for UAV signals as described above.

[0019] This application also provides a computer program product, including a computer program that, when executed by a processor, implements an interference identification method for UAV signals as described above.

[0020] The interference identification method for UAV signals provided in this application involves: acquiring the interfered UAV signal; performing short-time Fourier transform on the UAV signal to obtain a two-dimensional time-frequency map corresponding to the UAV signal; inputting the two-dimensional time-frequency map into an interference category identification model, which outputs the interference type to which the interference signal belongs; and inputting the two-dimensional time-frequency map into an interference target detection model based on the parallel fusion of convolution operation path and self-attention operation path, which outputs the prediction box and confidence score corresponding to the interference signal. The interference signal identification result includes the interference type, prediction box, and confidence score. The prediction box indicates the spatiotemporal location of the interference signal in the two-dimensional time-frequency map, and the confidence score indicates the confidence level of the prediction box. When performing interference identification on UAV signals, a short-time Fourier transform is first used to convert and process the UAV signal, enabling joint time-frequency domain interference feature extraction and obtaining a two-dimensional time-frequency map corresponding to the UAV signal. Based on the two-dimensional time-frequency map and a neural network-based interference category identification model, the interference type of the interference signal can be accurately determined. Furthermore, the two-dimensional time-frequency map is input into an interference target detection model based on the parallel fusion of convolution operation paths and self-attention operation paths. This transforms the estimation of key parameters such as the frequency band, bandwidth, and period of the interference signal into a computer vision target detection task based on the two-dimensional time-frequency map, achieving accurate identification of interference types and precise localization of key parameters, thereby effectively improving the accuracy of interference identification. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in 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, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the architecture of an interference identification system provided in an embodiment of this application.

[0023] Figure 2 This is a flowchart illustrating a method for identifying interference in drone signals, as provided in an embodiment of this application.

[0024] Figure 3 This is a schematic diagram of the structure of an interference target detection model provided in an embodiment of this application.

[0025] Figure 4 This is a schematic flowchart illustrating a method for outputting a prediction box and a confidence score corresponding to an interference signal through an interference target detection model, as provided in an embodiment of this application.

[0026] Figure 5This is a schematic diagram of the structure of a backbone network provided in an embodiment of this application.

[0027] Figure 6 This is a schematic diagram of the structure of the first dynamic convolution and self-attention bottleneck module provided in the embodiments of this application.

[0028] Figure 7 This is a schematic diagram of the structure of an AIFI module provided in an embodiment of this application.

[0029] Figure 8 This is a schematic diagram of a CFFM structure provided in an embodiment of this application.

[0030] Figure 9 Experimental curves showing the estimation of key parameters for the four detection methods provided in this application under different signal-to-noise ratios.

[0031] Figure 10 Bar chart showing resource consumption and real-time performance indicators of different methods or models in actual experiments provided in the embodiments of this application.

[0032] Figure 11 This is a schematic diagram of the structure of an interference identification device for drone signals provided in an embodiment of this application.

[0033] Figure 12 This is a schematic diagram of the physical structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] In the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone, where A and B can be singular or plural. In the textual description of this application, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0036] The technical solution provided in this application can be applied to scenarios involving interference identification of drone signals. Examples include drone countermeasures and security monitoring, drone anti-interference communication and link adaptation, and interference source localization and electronic warfare.

[0037] When performing drone signal recognition, deep learning-based recognition methods are usually used. Although these methods can improve the automation of feature extraction to some extent, their model structures are mostly designed for a single dimension. In complex interference scenarios, they can only determine whether the interference signal exists or not, but cannot identify the precise time point, frequency range, and other spatiotemporal parameters of the interference, resulting in low accuracy of interference recognition.

[0038] Based on this, embodiments of this application provide a method for identifying interference in UAV signals, which involves acquiring the interfered UAV signal; performing short-time Fourier transform on the UAV signal to obtain a two-dimensional time-frequency map corresponding to the UAV signal; inputting the two-dimensional time-frequency map into an interference category identification model, which outputs the interference type to which the interference signal belongs; and inputting the two-dimensional time-frequency map into an interference target detection model based on the parallel fusion of convolution operation path and self-attention operation path, which outputs the prediction box and confidence score corresponding to the interference signal. The identification result of the interference signal includes the interference type, prediction box, and confidence score. The prediction box indicates the spatiotemporal position of the interference signal in the two-dimensional time-frequency map, and the confidence score indicates the confidence level of the prediction box.

[0039] When performing interference identification on UAV signals, a short-time Fourier transform is first used to convert and process the UAV signal, enabling joint time-frequency domain interference feature extraction and obtaining a two-dimensional time-frequency map corresponding to the UAV signal. Based on the two-dimensional time-frequency map and a neural network-based interference category identification model, the interference type of the interference signal can be accurately determined. Furthermore, the two-dimensional time-frequency map is input into an interference target detection model based on the parallel fusion of convolution operation paths and self-attention operation paths. This transforms the estimation of key parameters such as the frequency band, bandwidth, and period of the interference signal into a computer vision target detection task based on the two-dimensional time-frequency map, achieving accurate identification of interference types and precise localization of key parameters, thereby effectively improving the accuracy of interference identification.

[0040] To train the interference category identification model and interference target detection model involved in the embodiments of this application, and to improve the real-time performance of interference detection, an interference identification system can be built in this embodiment. This system includes a drone device, an interference signal generator, a Universal Software Radio Peripheral (USRP) receiver, and an analysis terminal. The interference signal generator generates interference noise, and the receiver collects radio frequency signal data of the drone under various interference signals and sends it to the analysis terminal. For example, see [link to example]. Figure 1 As shown, Figure 1This is a schematic diagram of the architecture of an interference identification system provided in an embodiment of this application. The UAV device and the remote controller communicate normally via radio frequency (RF) signals, such as frequency hopping signals or orthogonal frequency division multiplexing (OFDM) signals. Non-cooperative parties can set various interference modes through an interference signal generator, radiating interference noise into space via a transmitting antenna, thus interfering with the communication link. RF signals in the environment, including normal communication signals and interference signals, are intercepted by a signal receiver and sent to the storage repository of an analysis terminal for storage and management, while simultaneously recording metadata such as acquisition time and interference source parameters. Subsequently, computer software preprocesses the acquired RF signals, including signal segmentation, data normalization, and short-time Fourier transform. The Transform (STFT) converts a one-dimensional radio frequency signal into a two-dimensional time-frequency map. The preprocessed two-dimensional time-frequency map can be used to train deep learning models, including interference category recognition models and interference target detection models. This allows the trained interference category recognition models and interference target detection models to be deployed on drones in flight or ground stations to achieve accurate identification of interference types and precise positioning of key parameters.

[0041] It is understood that the execution subject of the drone signal interference identification method provided in this application can be a computer, a server, or a specially set drone signal interference identification device, or a drone signal interference identification device set in the electronic device. The drone signal interference identification device can be implemented by software, hardware, or a combination of both, and can be set according to actual needs.

[0042] The interference identification method for UAV signals provided in this application will be described in detail below through several specific embodiments. It is understood that these specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0043] Figure 2 This is a flowchart illustrating a method for identifying interference in UAV signals provided in an embodiment of this application. For example, please refer to... Figure 2 As shown, the interference identification method for the drone signal may include: S201. Acquire the jammed drone signal.

[0044] For example, in the embodiments of this application, when acquiring the interfered drone signal, it can be obtained through field measurement, that is, by deploying drone equipment, interference signal generator and signal receiver in a real or simulated environment, applying interference and intercepting drone signal during actual flight; or it can be obtained by using hardware-in-the-loop simulation, using a vector interference signal generator to generate drone communication signal and interference signal, superimposing them through a power divider or combiner and then inputting them into a signal receiver to obtain the interfered drone signal, etc. The specific settings can be configured according to actual needs.

[0045] S202. The UAV signal is converted and processed using short-time Fourier transform to obtain a two-dimensional time-frequency diagram corresponding to the UAV signal.

[0046] Among them, the two-dimensional time-frequency diagram contains the time-frequency domain information of the interference signal, which can be used to reflect the energy distribution of the interference signal in the time-frequency domain.

[0047] For example, in an embodiment of this application, the short-time Fourier transform is used to convert and process the UAV signal to obtain a two-dimensional time-frequency diagram corresponding to the UAV signal, which may include: A time-localized time window function is applied to the UAV signal, and Fourier transform calculations are performed on the signal segments within the current time window to obtain the spectral information of the signal segments corresponding to the current time window. The current time window is then slid along the time axis to obtain the spectral maps of the signal segments corresponding to multiple time windows. These spectral maps constitute a two-dimensional time-frequency diagram. There is overlap between adjacent time windows, thus introducing a short-time Fourier transform. Local spectral analysis of the signal is performed through the time window function and the sliding window function. As the window slides along the time axis, a series of spectral maps corresponding to different time points are obtained, forming a two-dimensional time-frequency representation. This method can intuitively and completely preserve the dynamic evolution characteristics of the interference signal in the time and frequency dimensions, providing a richer and more discriminative input representation for subsequent high-precision identification and parameter estimation, thereby better capturing subtle changes in the UAV signal. Its continuous form is defined as: in, Indicates the time of the drone signal Nearby, frequency The spectral content at a given point is the mathematical expression of a two-dimensional time-frequency graph. Indicates time The changing one-dimensional time-domain signal, i.e., the received jammed UAV signal, Indicated by time A localized window centered on the subject, used for capturing Spectral information of nearby signal segments, Represents the time variable of integration. Represents the complex exponential kernel, a basis function of the Fourier transform, used to detect frequencies of 1 in a signal. The ingredients, It represents the imaginary unit.

[0048] After discretizing the above continuous form, its discretized form can be seen in the following formula: in, Represents the discrete STFT coefficients, indicating the UAV signal at the th... The first time window, the first The spectral content at each frequency point is the specific numerical expression of the two-dimensional time-frequency graph during computer processing, corresponding to the continuous form mentioned above. 'm' represents the sample index within the window, and 'm' represents the sampling point position number within the current window, corresponding to the integral time variable in the continuous form. The value range is 0 to , This represents the window length, i.e., the number of sampling points contained in each time window, which determines the frequency resolution and corresponds to the effective duration of the window function in the continuous form. This refers to discrete UAV signals, i.e., received signals from interfered UAVs. Discrete sequence after sampling Indicates drone signal The Middle Sampling points correspond to UAV signals in continuous form. , Let represent a discrete window function, a weighted sequence of coefficients for each sampling point within the window, with a length of . , corresponding to the window function in the continuous form , Represents the discrete complex exponential kernel, the discrete form of the basis functions of the Fourier transform, used to detect the first complex exponential kernel in a signal. Each frequency component corresponds to a complex exponential kernel in the continuous form. .

[0049] It is understood that in this embodiment of the application, when performing local spectral analysis of the signal by introducing a sliding window function, in order to improve the discontinuity between signal segments caused by window division, an overlapping window method is adopted. This can effectively improve the smoothness and continuity of the window division result, and the number of overlapping points can be adjusted as a preset parameter. When there is overlap between adjacent windows, a finer time resolution and a more coherent time-frequency representation can be obtained, which helps to capture subtle changes in the UAV signal. In this way, the time-frequency representation of the UAV signal can be obtained without complex mathematical processing. By reasonably setting the window size and sliding step size, the latency in the real-time processing can be effectively controlled, ensuring that information is acquired in a timely manner while maintaining low latency. This allows for rapid response to emergencies in the UAV communication process, thereby ensuring the stability and reliability of the communication link to a certain extent.

[0050] For example, the time window function can be one of a Hamming window, a Hanning window, or a Gaussian window, and the overlap rate between two adjacent time windows can be 50% to 75%, which can be set according to actual needs.

[0051] For example, when the spectrum diagrams of the signal segments corresponding to multiple time windows are used to form a two-dimensional time-frequency diagram, the spectrum diagrams of the signal segments corresponding to multiple time windows can be merged by splicing or superimposing to form a two-dimensional time-frequency diagram corresponding to the UAV signal. The specific settings can be configured according to actual needs.

[0052] 203. Input the two-dimensional time-frequency image into the interference category identification model, and output the interference type to which the interference signal belongs through the interference category identification model; input the two-dimensional time-frequency image into the interference target detection model based on the parallel fusion of convolution operation path and self-attention operation path, and output the prediction box and confidence score corresponding to the interference signal through the interference target detection model.

[0053] The identification results of the interference signal include the interference type, prediction box, and confidence score. The prediction box is used to indicate the spatiotemporal location of the interference signal in the two-dimensional time-frequency diagram, and the confidence score is used to indicate the confidence of the prediction box.

[0054] For example, the interference category recognition model can be a ResNet-based network model. ResNet is a deep neural network that introduces residual blocks, allowing the input of a layer to skip several layers and be passed directly to subsequent layers. This "skip connection" setting enables gradients to be propagated more smoothly back to shallower layers during backpropagation, effectively alleviating the gradient vanishing problem in deep networks. This allows for training very deep networks, improving the recognition performance of interference signals, and extracting the features needed for subsequent steps.

[0055] For example, ResNet can be one of ResNet50, ResNet101, or ResNet152, which can be set according to actual needs.

[0056] Taking the interference category recognition model as an example, based on the ResNet101 network model, the training set consists of 16,128 two-dimensional UAV signal time-frequency images, and the test set consists of 2,688 two-dimensional UAV signal time-frequency images, both generated using MATLAB software. The loss function used during training is Cross Entropy Loss (CEL), and the optimizer is Stochastic Gradient Descent (SGD). The base learning rate is 3e-4, combined with a cosine annealing strategy, and the weight decay coefficient is set to 0.001. The batch size is set to 16, and the total number of training epochs is set to 25 to train the interference category recognition model. See Table 1 below for an example. Table 1 shows the architecture configuration information of an interference category recognition model provided in this embodiment. This interference category recognition model can achieve higher recognition accuracy and faster inference speed, thereby accurately determining the interference type of the interference signal.

[0057] Table 1

[0058] Based on Table 1 above, the convolutional layers and max-pooling layers in Table 1 can be used as shallow feature extraction layers, residual block ×3, residual block ×4, residual block ×23, and residual block ×3 can be used as deep semantic extraction layers, global pooling can be used as feature compression layers, and fully connected layers can be used as classification output layers. The input to the interference category recognition model is a two-dimensional time-frequency image. The shallow feature extraction layer is used to perform preliminary visual feature extraction on the input two-dimensional time-frequency image, and the convolutional layers use a relatively large 7×7×7×7 matrix. Convolutional kernels capture low-level features such as edges, textures, and energy patches in the 2D time-frequency image. Max pooling layers reduce resolution through downsampling, enhancing the translation invariance of features while preserving the main feature responses. The deep semantic extraction layer, as the backbone of ResNet101, transforms spatial features into semantic features by progressively increasing depth and channel number. Each residual block alleviates the gradient vanishing problem through skip connections, enabling ResNet101 to be trained to a depth of 101 layers, thereby extracting richer time-frequency features. The feature compression layer transforms spatial features into global features, adapting them to the input of the fully connected layer. The classification output layer acts as a classifier, mapping the 2048-dimensional feature vector output by the global average pooling layer to six specific interference categories, such as single-tone interference, frequency sweep interference, impulse interference, broadband noise interference, and narrowband noise interference. The fully connected layer is usually followed by a Softmax activation function, which outputs the probability of each interference category, ultimately determining the type of interference the signal belongs to.

[0059] For example, in the embodiments of this application, the interference type may include at least one of single-tone interference, frequency sweep interference, pulse interference, broadband noise interference, and narrowband noise interference, which can be set according to actual needs.

[0060] For example, in the embodiments of this application, the prediction box can be represented by the coordinates of the center point and the width and height, or by the coordinates of the upper left corner and the lower right corner, which can be set according to actual needs.

[0061] In this embodiment, on the one hand, an interference category identification model is used to achieve high-accuracy interference type classification; on the other hand, an interference target detection model is used to transform the parameter estimation problem into a target detection task in a two-dimensional time-frequency graph. The joint optimization mechanism realizes accurate identification of interference types and precise positioning of key parameters, thereby effectively improving the accuracy of interference identification.

[0062] As can be seen, in this embodiment, the interference of the UAV signal is acquired; the UAV signal is converted and processed using short-time Fourier transform to obtain a two-dimensional time-frequency map corresponding to the UAV signal; the two-dimensional time-frequency map is input into the interference category identification model, and the interference category identification model outputs the interference type to which the interference signal belongs; the two-dimensional time-frequency map is input into the interference target detection model based on the parallel fusion of convolution operation path and self-attention operation path, and the interference target detection model outputs the prediction box and confidence score corresponding to the interference signal; wherein, the identification result of the interference signal includes the interference type, prediction box and confidence score, the prediction box is used to indicate the spatiotemporal position of the interference signal in the two-dimensional time-frequency map, and the confidence score is used to indicate the confidence of the prediction box. When performing interference identification on UAV signals, a short-time Fourier transform is first used to convert and process the UAV signal, enabling joint time-frequency domain interference feature extraction and obtaining a two-dimensional time-frequency map corresponding to the UAV signal. Based on the two-dimensional time-frequency map and a neural network-based interference category identification model, the interference type of the interference signal can be accurately determined. Furthermore, the two-dimensional time-frequency map is input into an interference target detection model based on the parallel fusion of convolution operation paths and self-attention operation paths. This transforms the estimation of key parameters such as the frequency band, bandwidth, and period of the interference signal into a computer vision target detection task based on the two-dimensional time-frequency map, achieving accurate identification of interference types and precise localization of key parameters, thereby effectively improving the accuracy of interference identification.

[0063] Based on the above Figure 2 The illustrated embodiment, for example, in this application embodiment, the interference target detection model may include a backbone network and a feature processing module. For example, see [link to relevant documentation]. Figure 3 As shown, Figure 3 This is a schematic diagram of the structure of an interference target detection model provided in an embodiment of this application. In S203 above, the two-dimensional time-frequency map is input into the interference target detection model based on the parallel fusion of convolution operation path and self-attention operation path. The specific implementation of the interference target detection model outputting the prediction box and confidence score corresponding to the interference signal can be found below. Figure 4 The example shown.

[0064] Figure 4 This application provides a flowchart illustrating a method for outputting a prediction box and confidence score corresponding to an interference signal through an interference target detection model. For example, it can be combined with... Figure 4 As shown, the method may include: S401. Input the two-dimensional time-frequency map into the backbone network based on the parallel fusion of convolution operation path and self-attention operation path. Perform multi-scale convolution on the two-dimensional time-frequency map through the backbone network and output the target multi-scale feature map corresponding to the two-dimensional time-frequency map.

[0065] For example, in an embodiment of this application, the backbone network may include a first high-dimensional feature extraction module and a second high-dimensional feature extraction module based on the parallel fusion of convolutional operation paths and self-attention operation paths. For example, see [link to relevant documentation]. Figure 5 As shown, Figure 5 This is a schematic diagram of a backbone network provided in an embodiment of this application. A two-dimensional time-frequency image is input into a backbone network based on the parallel fusion of convolutional operation paths and self-attention operation paths. The backbone network performs multi-scale convolutions on the two-dimensional time-frequency image, outputting a target multi-scale feature map corresponding to the two-dimensional time-frequency image. This may include: The two-dimensional time-frequency image is input into the first high-dimensional feature extraction module, which extracts high-dimensional features from the image and outputs the high-dimensional features. These high-dimensional features are then input into the second high-dimensional feature extraction module, which performs parallel feature extraction based on convolutional and self-attention operation paths. The extracted features are then fused to output a multi-scale feature map of the target. This two-stage backbone network design, involving "coarse extraction followed by fine processing," provides high-quality input features for subsequent AIFI, CFFM, and decoders, thereby supporting the entire interference target detection model to achieve high-precision and real-time recognition performance.

[0066] For example, in an embodiment of this application, the second high-dimensional feature extraction module may include multiple dynamically convolutional and self-attention bottleneck modules connected in series. For example, it may combine... Figure 5 As shown, where, Figure 5 Each `B0ttleNeck_DynaCoAtt` in the module can be denoted as a dynamic convolution and self-attention bottleneck module. High-dimensional features are input into a second high-dimensional feature extraction module that uses a parallel fusion of convolutional and self-attention operation paths. This second module performs parallel feature extraction based on convolutional and self-attention operation paths and then fuses the extracted features to output a target multi-scale feature map, which may include: High-dimensional features are input into the first dynamic convolution and self-attention bottleneck module. This module performs parallel feature extraction based on convolution and self-attention operation paths, and then fuses the extracted features to output a first multi-scale feature map. For the other dynamic convolution and self-attention bottleneck modules (excluding the first one), these modules perform parallel feature extraction based on convolution and self-attention operation paths on the second multi-scale feature map output by the previous module, and then fuse the extracted features to output a third multi-scale feature map. The third multi-scale feature map output by the last dynamic convolution and self-attention bottleneck module is the target multi-scale feature map. By sequentially connecting multiple dynamic convolutions and self-attention bottleneck modules, the backbone network achieves progressive refinement of local and global features, hierarchical expansion of the receptive field, and layer-by-layer enrichment of multi-scale features, while maintaining good training stability and modular flexibility. This enables the final output multi-scale feature map of the target to comprehensively and with high quality represent the time-frequency characteristics of the interference signal, laying a solid foundation for the subsequent AIFI, CFFM modules and the final interference target detection.

[0067] It should be noted that, in the embodiments of this application, the structures of each of the above-mentioned multiple dynamically convolutional and self-attention bottleneck modules connected in sequence are similar. For example, in the embodiments of this application, taking the first dynamically convolutional and self-attention bottleneck module as an example, the first dynamically convolutional and self-attention bottleneck module may include a convolutional projection unit, a convolutional unit based on the convolutional operation path, and a self-attention unit based on the self-attention operation path. For example, see [link to relevant documentation]. Figure 6 As shown, Figure 6 This embodiment of the application provides a schematic diagram of the structure of a first dynamic convolution and self-attention bottleneck module. High-dimensional features are input into this module, which performs parallel feature extraction based on convolution and self-attention operation paths. The extracted features are then fused to output a first multi-scale feature map, which may include: High-dimensional features are input into the convolutional projection unit, which performs linear projection on the high-dimensional features, outputting query Q, key K, and value V. Query Q, key K, and value V are then input into the convolutional unit and the self-attention unit, respectively. The convolutional unit performs depthwise separable convolution on query Q, key K, and value V, outputting a convolutional feature map. The self-attention unit performs multi-head self-attention on query Q, key K, and value V, outputting a self-attention feature map. The convolutional and self-attention feature maps are then fused to obtain the first multi-scale feature map. By uniformly generating query Q, key K, and value V through the convolutional projection unit, and then feeding them into the convolutional and self-attention units for parallel processing and fusion, the first dynamic convolution and self-attention bottleneck module achieves multiple advantages: high computational efficiency, rich features, lightweight parameters, and stable training. This allows the output first multi-scale feature map to simultaneously contain fine local information and broad global context, laying the foundation for the subsequent stacking of multiple modules.

[0068] The number of channels for query Q, key K, and value V can be the same or different, and can be set according to actual needs.

[0069] For example, the kernel size for depthwise separable convolution operations can be 3×3 or 5×5, and the number of heads for multi-head self-attention operations can be 8 or 16.

[0070] For example, in an embodiment of this application, fusing the convolutional feature map and the self-attention feature map to obtain a first multi-scale feature map may include: The gating coefficients corresponding to the convolutional unit and the self-attention unit are determined respectively. Based on the gating coefficients corresponding to the convolutional unit and the self-attention unit, the convolutional feature map and the self-attention feature map are weighted and fused to obtain the first multi-scale feature map.

[0071] The gating coefficients are learned from the input features through a fully connected layer, or they may be trainable scalar parameters. and satisfy ,or and Normalize to the (0, 1) interval using the Sigmoid function.

[0072] For example, in this embodiment of the application, the interference target detection model can be a reproduction of the Real-Time Detection Transformer (RT-DETR) architecture, replacing the original backbone network with ResNet50, and introducing the aforementioned [structure] at the BottleNeck structure position in the ResNet50 network. Figure 4The second high-dimensional feature extraction module shown replaces the original standard convolutional layer, corresponding to layers P3 to P5 of feature extraction. It performs parallel feature extraction of high-dimensional features based on convolutional operation path and self-attention operation path, and fuses the parallel extracted features to output a target multi-scale feature map.

[0073] It is worth noting that in the aforementioned second high-dimensional feature extraction module, each dynamic convolution and self-attention bottleneck module, i.e. Figure 5 The B0ttleNeck_DynaCoAtt algorithm employs a dynamic weight allocation mechanism. Firstly, it discovers that convolution and self-attention can share a 1×1 projection in the first stage to avoid redundant computation and reduce the number of parameters and computational burden. Secondly, it designs a dual-path feature reuse mechanism, decoupling local feature extraction and global relationship modeling into two parallel paths: a convolution operation path and a self-attention operation path, which perform subsequent spatial aggregation operations separately. (See above for details.) Figure 6 As shown, where, and These represent the height and width of the input high-dimensional feature, respectively. This represents the number of channels in the input high-dimensional feature. This indicates the kernel size of the convolutional layer. and This represents the gating coefficients corresponding to the convolutional unit and the self-attention unit, respectively. Research shows that convolution and self-attention decomposition can include the following two stages: One stage involves convolution decomposition, as shown in the following formula: in, Indicates through convolution kernel The intermediate features obtained after projecting onto position (i, j) This represents the weight matrix corresponding to the relative offset (p,q) in the standard convolutional kernel, responsible for projecting the input features into the output channel space. This represents the feature vector at spatial location (i, j) of the input feature map, with dimension Cin = the number of channels. This indicates that after the convolution operation, the feature vector at position (i, j) on the output feature map is generated. This indicates a shift operation, which shifts the intermediate feature relative to its position. Move to the target location to achieve spatial aggregation.

[0074] Another stage is self-attention decomposition, which can be seen in the following formula: in, , and Indicates the first The query, key, and value vector generated by an attention head at position (i, j) , and Indicates the first The projection matrix of the query, key, and value corresponding to each attention head. A represents the total number of self-attention heads. Indicates attention weights. This indicates that after self-attention, the feature vector at position (i, j) on the output feature map is generated.

[0075] The above decomposition and calculation process, along with the integration of commonalities, involves Convolution projection, convolution kernel Corresponding to self-attention , , The computational complexity is concentrated in the first stage, satisfying The second stage involves lightweight spatial aggregation operations. Based on this, in this embodiment, a unified convolution projection unit is designed to decompose the standard convolution kernel into multiple... Convolutional kernels, each processing a different subset of features, also transform the generation of queries, keys, and values ​​in the self-attention mechanism into... Convolution operations reduce computational burden.

[0076] Next, in the process After the convolution operation, the query Q, key K, and value V are input into the convolution unit and the self-attention unit, respectively.

[0077] The convolutional unit performs depthwise separable convolution operations on the query Q, key K, and value V, including: in, This represents the convolutional feature map output by the convolutional unit. This indicates the kernel size, k represents the index of a position within the kernel, and DepthwiseConv represents the depthwise separable convolution operation. This represents the k-th shift weight tensor. Its implementation is as follows: in, This represents the weight values ​​of the depthwise convolution kernel at position (p,q) and channel cc. This represents the learnable weight parameters. This represents the Kronecker delta, which is 1 when c = k and 0 otherwise. Denotes the indicator function of set K, when The value is 1 if the condition is met, and 0 otherwise. The constraint is: The self-attention unit performs multi-head self-attention operations on query Q, key K, and value V, including: in, This represents the self-attention feature map output by the self-attention unit. This represents the dot product matrix of the query and the key, and calculates the similarity between all positions. This indicates that row-wise normalization yields the attention weight matrix.

[0078] Finally, the convolutional feature maps output by the convolutional units and the self-attention feature maps output by the self-attention units are fused, as shown in the following formula, to obtain the first multi-scale feature map. In this process, the convolutional part ensures that the model can accurately extract and utilize local detail information, while the self-attention part provides a global horizon, enabling the model to integrate information over a larger scope and promote the flow and interaction of information at different scales.

[0079] in, This represents the first multi-scale feature map obtained by weighted fusion of the convolutional feature map and the self-attention feature map. and These represent the gating coefficients corresponding to the convolutional unit and the self-attention unit, respectively, which can adapt to different task requirements, such as the averaging in interference signal classification tasks. In complex environments, parameter estimation is used in target detection tasks. The specific settings can be configured according to actual needs. Furthermore, in this embodiment, considering only the use of... Convolutional layers process the input feature map to obtain the Q, K, and V matrices, then perform self-attention transformations. This results in some loss in feature extraction. Therefore, the concatenation after attention is not simply a concatenation of the attention heads; it is also multiplied by a trainable output matrix with weight parameters. This serves as compensation for feature extraction at the spatial level of each channel. By introducing a channel attention and relative position encoding mechanism based on convolution and attention, the ability of the interference target detection model to perceive local details and global structures in two-dimensional time-frequency images is significantly enhanced. While ensuring real-time performance, it greatly improves the detection and localization accuracy of dense, overlapping, or weak-energy interference signals.

[0080] S402. Input the multi-scale features of the target into the feature processing module. The feature processing module performs feature interaction and fusion on the multi-scale features of the target and outputs the prediction box and confidence score corresponding to the interference signal.

[0081] For example, in the embodiments of this application, the feature processing module includes an attention-based intrascale feature interaction (AIFI) module, a cross-scale feature fusion module (CFFM), and a decoding module, as described above. Figure 3 As shown, the multi-scale features of the target are input into the feature processing module. The feature processing module performs feature interaction and fusion on the multi-scale features of the target, and outputs the prediction box and confidence score corresponding to the interference signal. This may include: The target multi-scale feature map is input into the AIFI module, which performs intra-scale feature self-attention fusion to output the first fused feature. Both the target multi-scale feature map and the first fused feature are input into CFFM, where they are fused to obtain the second fused feature. This second fused feature is then input into the decoding module, which decodes it based on decoupled location and content queries, outputting the predicted bounding box and confidence score corresponding to the interference signal. Through this progressive processing of intra-scale interaction, cross-scale fusion, and decoupled prediction, the feature processing module achieves a complete chain of global context enhancement, multi-scale information fusion, and accurate target decoding. This ensures that the final output predicted bounding box and confidence score are both accurate and reliable, providing high-quality feature assurance for real-time detection of UAV interference signals.

[0082] For example, in an embodiment of this application, decoding the second fused feature based on the decoupled location and content query pair by the decoding module, and outputting the confidence score corresponding to the interference signal, may include: The second fused feature is decoded based on the decoupled location and content queries, outputting an initial confidence score. The final confidence score is then determined based on this initial confidence score and the IOU-predicted confidence score. This decoupled query approach ensures accurate extraction of location and category features, while IoU awareness ensures a strong correlation between the score and location quality. This not only improves the accuracy but also the reliability of the confidence score, making it particularly suitable for real-time scenarios requiring rapid response. For example, in an embodiment of this application, determining a confidence score based on an initial confidence score and an IOU confidence score predicted based on IOU includes: The product of the initial confidence score and the IoU confidence score is determined; this product is then used as the confidence score. This method of multiplying the initial confidence score by the IoU confidence score to obtain the final confidence score essentially achieves deep coupling between the classification and localization tasks through mathematical multiplication. This design ensures that the final score truly reflects the overall quality of the predicted bounding boxes, thereby significantly improving the accuracy and reliability of UAV interference signal detection.

[0083] For example, see Figure 7 As shown, Figure 7 This is a schematic diagram of the structure of an AIFI module provided in an embodiment of this application. The AIFI module performs intra-scale feature self-attention fusion on the multi-scale feature map of the target, as shown in the following formula: in, This refers to the target multi-scale features output by the backbone network based on the parallel fusion of convolutional operation paths and self-attention operation paths. ( () indicates the flattening operation. Indicates attention weights. Represents the query vector. Represents the key vector. Representing feature dimension, Represents the sequence length, exp( () represents the exponential function, converting the dot product to a non-negative value. For the projection matrix, Representing multi-scale feature maps The feature vector at position i, Indicates the output features of Transformer The feature vector at position j, This represents the first fusion feature output by the AIFI module.

[0084] The AIFI module uses a Transformer encoder to extract high-dimensional semantic features. It consists of stacked identical layers, each containing two core sub-layers: one is a multi-head self-attention mechanism: given an input sequence The self-attention mechanism uses three learnable matrices Generate queries, keys, and values; note the scaling dot product calculation: Multi-head attention is obtained by concatenating multiple attention heads and then performing a linear transformation. Each head's output contains a different attention allocation, as calculated above. The second sub-layer consists of a feedforward network (FFN) following each attention sub-layer, enabling the learning and capture of more non-linear features about the time-frequency image of the signal. in, This represents the vector at each position in the sequence after the input feature vector has undergone self-attention processing. These represent the weight matrix and bias of the first fully connected layer, respectively. This represents the weight matrix and bias of the second fully connected layer. , , The feature dimensions representing the input and output, This represents the dimension of the hidden layers in a feedforward network, typically greater than [value missing]. , This represents the output of the feedforward network, with the same shape as the input x. Each sublayer uses residual connections and layer normalization. in, This represents the output of the sublayer. Indicates residual connection, ( This indicates layer normalization.

[0085] The target multi-scale features output by the backbone network and the first fused features output by the AIFI module are both input into CFFM. For example, see [link to example]. Figure 8 As shown, Figure 8 This is a schematic diagram of the structure of CFFM provided in an embodiment of this application. CFFM performs feature fusion on the target multi-scale feature map and the first fusion feature. The channel attention module is used to weight the feature map at each scale, and then the spatial attention module is used to further process the weighted feature map to integrate detailed features and contextual information. The aim is to enhance the network's tolerance to scale differences and improve its ability to perceive subtle targets, thereby obtaining the second fusion feature.

[0086] For example, the feature fusion equation is defined as follows: in, This represents the intermediate fusion features, specifically the feature map at the current scale (l) after initial stitching and fusion. Indicates feature splicing, This represents the feature map at the k-th scale, derived from the output of the backbone network or the AIFI module. This represents a spatial gating weight map, where the value at each spatial location indicates the proportion that location should be biased towards the fused features. ( ) represents the activation function. This represents the final output feature at the current scale l. For element-wise multiplication, the gating coefficient Adaptive feature selection is achieved. The interference identification process using the interference category recognition model in this application can be formalized as follows: in, Represents a two-dimensional time-frequency graph. This represents the multi-scale features of the target output by the backbone network. This represents the first fusion feature output by the AIFI module. This represents the second fusion feature of the CFFM output. This represents the prediction result output by the decoder, including information such as the prediction bounding box and confidence score. This indicates a decoupled location and content query pair.

[0087] For example, in an embodiment of this application, in order to further improve the accuracy of the feasibility score, the predicted IOU confidence score can be combined to jointly determine the confidence score corresponding to the interference signal, as shown in the following formula: in, This represents the confidence score corresponding to the interference signal. This indicates that the initial confidence score is output by decoding the second fused feature based on the decoupled location and content query pair. This represents the predicted IOU confidence score. By combining the predicted IOU confidence score with the corresponding confidence score of the interference signal, the ability to identify high IoU features can be enhanced, while the response to low IoU features can be suppressed. This achieves consistency between classification score and localization quality, resulting in both high classification score and high IoU score.

[0088] Figure 9The figures show experimental curves of key parameter estimation for the four detection methods provided in this application under different signal-to-noise ratios. The horizontal axis represents the signal-to-noise ratio (SNR), ranging from -10 dB to 25 dB; the vertical axis represents the relative error of bandwidth estimation. The figures also show the bandwidth estimation error curves of the interference target detection model provided in this application compared with three other methods, including traditional methods based on handcrafted features, YOLO11m-SEG, and RT-DETR, two mainstream deep neural network methods, under different signal-to-noise ratios.

[0089] Experimental results show that, under all signal-to-noise ratio (SNR) conditions, the relative error of bandwidth estimation obtained by the interference target detection model provided in this application is lower than that of the comparative methods. Especially in the low SNR region (-10 dB to 0 dB), the error curve obtained by the interference target detection model provided in this application is significantly lower than that of other methods, demonstrating stronger noise resistance and parameter estimation accuracy. As the SNR increases, the estimation errors of each comparative method gradually decrease, but the error obtained by the interference target detection model provided in this application remains at the lowest level, verifying its superiority and robustness in the task of interference signal parameter estimation.

[0090] Figure 10 The bar chart shows the resource consumption and real-time performance indicators of different methods or models in the actual test provided in the embodiments of this application. The chart compares the performance of traditional methods, RT-DETR models and interference target detection models provided in this application in terms of the highest memory usage, inference latency and inference frames per second in three key performance indicators.

[0091] In terms of maximum video memory usage, traditional methods use approximately 500 MB, RT-DETR uses approximately 2000 MB, while the interference target detection model provided in this application uses approximately 1500 MB. The results show that the interference target detection model provided in this application reduces video memory usage by approximately 25% compared to RT-DETR, demonstrating its optimization advantage in resource utilization.

[0092] In terms of inference latency, traditional methods have a latency of approximately 10 ms, RT-DETR has a latency of approximately 40 ms, while the interference target detection model provided in this application has a latency of approximately 30 ms. The inference latency of the interference target detection model provided in this application is 25% lower than that of RT-DETR, indicating that it achieves a faster response speed while maintaining high accuracy.

[0093] In terms of inference frames per second (FPS), traditional methods can reach 80 FPS, RT-DETR is about 60 FPS, and the interference target detection model provided in this application is about 50 FPS. Although the frame rate of the interference target detection model provided in this application is slightly lower than that of RT-DETR, it still maintains a high real-time processing capability, and combined with memory usage and latency metrics, it shows a better balance between resource efficiency and real-time performance.

[0094] In summary, the interference target detection model provided in this application outperforms RT-DETR in terms of memory usage and inference latency. Although the number of inference frames is slightly reduced, it is still within the acceptable range for real-time applications. This fully verifies the effectiveness of the interference target detection model provided in this application in identifying UAV communication signal interference in actual electromagnetic environments, and provides an important reference for the application of UAV communication anti-interference technology.

[0095] The interference identification device for UAV signals provided in this application is described below. The interference identification device for UAV signals described below can be referred to in correspondence with the interference identification method for UAV signals described above.

[0096] Figure 11 This is a schematic diagram of the structure of a drone signal interference identification device provided in an embodiment of this application. For example, please refer to [link to relevant documentation]. Figure 11 As shown, the interference identification device 110 for the drone signal may include: Acquisition unit 1101 is used to acquire the interfered UAV signal; The conversion unit 1102 is used to convert the UAV signal using short-time Fourier transform to obtain a two-dimensional time-frequency diagram corresponding to the UAV signal. The identification unit 1103 is used to input the two-dimensional time-frequency image into the interference category identification model, and output the interference type to which the interference signal belongs through the interference category identification model; and to input the two-dimensional time-frequency image into the interference target detection model based on the parallel fusion of convolution operation path and self-attention operation path, and output the prediction box and confidence score corresponding to the interference signal through the interference target detection model. The identification result of the interference signal includes the interference type, the prediction box, and the confidence score. The prediction box is used to indicate the spatiotemporal position of the interference signal in the two-dimensional time-frequency diagram, and the confidence score is used to indicate the confidence of the prediction box.

[0097] For example, in an embodiment of this application, the conversion unit 1102 is used to perform conversion processing on the UAV signal using short-time Fourier transform to obtain a two-dimensional time-frequency diagram corresponding to the UAV signal, including: A time-localized time window function is applied to the UAV signal, and a Fourier transform is performed on the signal segment within the current time window to obtain the spectral information of the signal segment corresponding to the current time window. The current time window is slid along the time axis to obtain the spectrum diagrams of the signal segments corresponding to each of the multiple time windows; wherein, the spectrum diagrams of the signal segments corresponding to each of the multiple time windows constitute the two-dimensional time-frequency diagram, and there is overlap between adjacent time windows.

[0098] For example, in an embodiment of this application, the interference target detection model includes a backbone network and a feature processing module; The recognition unit 130 is used to input the two-dimensional time-frequency map into an interference target detection model based on the parallel fusion of convolution operation paths and self-attention operation paths, and output the prediction box and confidence score corresponding to the interference signal through the interference target detection model, including: The two-dimensional time-frequency map is input into the backbone network based on the parallel fusion of convolution operation path and self-attention operation path. The backbone network performs multi-scale convolution on the two-dimensional time-frequency map and outputs the target multi-scale feature map corresponding to the two-dimensional time-frequency map. The target multi-scale features are input into the feature processing module, which performs feature interaction and fusion on the target multi-scale features and outputs the prediction box and confidence score corresponding to the interference signal.

[0099] For example, in an embodiment of this application, the backbone network includes a first high-dimensional feature extraction module and a second high-dimensional feature extraction module based on the parallel fusion of convolution operation paths and self-attention operation paths; The recognition unit 1103 is used to input the two-dimensional time-frequency map into the backbone network based on the parallel fusion of convolution operation paths and self-attention operation paths, and to perform multi-scale convolution on the two-dimensional time-frequency map through the backbone network to output the target multi-scale feature map corresponding to the two-dimensional time-frequency map, including: The two-dimensional time-frequency graph is input into the first high-dimensional feature extraction module, and the high-dimensional feature extraction module performs high-dimensional feature extraction on the two-dimensional time-frequency graph and outputs high-dimensional features. The high-dimensional features are input into the second high-dimensional feature extraction module, which performs parallel feature extraction based on convolution operation path and self-attention operation path, and fuses the parallel extracted features to output the target multi-scale feature map.

[0100] For example, in an embodiment of this application, the second high-dimensional feature extraction module includes multiple dynamically convolutional and self-attention bottleneck modules connected in series. The recognition unit 1103 is used to input the high-dimensional features into the second high-dimensional feature extraction module, which is based on the parallel fusion of convolutional operation paths and self-attention operation paths. The second high-dimensional feature extraction module performs parallel feature extraction on the high-dimensional features based on convolutional operation paths and self-attention operation paths, and fuses the parallel extracted features to output the target multi-scale feature map, including: The high-dimensional features are input into the first dynamic convolution and self-attention bottleneck module. The high-dimensional features are extracted in parallel based on the convolution operation path and the self-attention operation path through the first dynamic convolution and self-attention bottleneck module. The extracted features are then fused to output the first multi-scale feature map. For the other dynamic convolution and self-attention bottleneck modules besides the first one, parallel feature extraction based on convolution operation path and self-attention operation path is performed on the second multi-scale feature map output by the previous dynamic convolution and self-attention bottleneck module through the other dynamic convolution and self-attention bottleneck modules, and the parallel extracted features are fused to output a third multi-scale feature map; wherein, the third multi-scale feature map output by the last dynamic convolution and self-attention bottleneck module among the other dynamic convolution and self-attention bottleneck modules is the target multi-scale feature map.

[0101] For example, in an embodiment of this application, the first dynamic convolution and self-attention bottleneck module includes a convolution projection unit, a convolution unit based on the convolution operation path, and a self-attention unit based on the self-attention operation path. The recognition unit 1103 is used to input the high-dimensional features into the first dynamic convolution and self-attention bottleneck module, and to perform parallel feature extraction on the high-dimensional features based on the convolution operation path and the self-attention operation path through the first dynamic convolution and self-attention bottleneck module, and to fuse the parallel extracted features to output a first multi-scale feature map, including: The high-dimensional features are input into the convolutional projection unit, and the high-dimensional features are linearly projected through the convolutional projection unit to output the query, key, and value. The query, key, and value are respectively input into the convolutional unit and the self-attention unit. The convolutional unit performs a depthwise separable convolution operation on the query, key, and value, and outputs a convolutional feature map. The self-attention unit performs a multi-head self-attention operation on the query, key, and value, and outputs a self-attention feature map. The convolutional feature map and the self-attention feature map are fused to obtain the first multi-scale feature map.

[0102] For example, in an embodiment of this application, the recognition unit 1103 is used to fuse the convolutional feature map and the self-attention feature map to obtain the first multi-scale feature map, including: Determine the gating coefficients corresponding to the convolutional unit and the self-attention unit respectively; Based on the gating coefficients corresponding to the convolutional unit and the self-attention unit, the convolutional feature map and the self-attention feature map are weighted and fused to obtain the first multi-scale feature map.

[0103] For example, in the embodiments of this application, the feature processing module includes an attention-based intra-scale feature interaction (AIFI) module, a cross-scale feature fusion (CFFM) module, and a decoding module; The identification unit 1103 is used to input the multi-scale features of the target into the feature processing module, perform feature interaction and fusion on the multi-scale features of the target through the feature processing module, and output the prediction box and confidence score corresponding to the interference signal, including: The target multi-scale feature map is input into the AIFI module, and the AIFI module performs intra-scale feature self-attention fusion on the target multi-scale feature map to output the first fused feature. Both the target multi-scale feature map and the first fused feature are input into the CFFM. The CFFM is used to fuse the target multi-scale feature map and the first fused feature to obtain the second fused feature. The second fused feature is input into the decoding module, which decodes the second fused feature based on decoupled location and content queries, and outputs the prediction box and confidence score corresponding to the interference signal.

[0104] For example, in an embodiment of this application, the identification unit 1103 is used to decode the second fused feature based on the decoupled location and content query by the decoding module, and output the confidence score corresponding to the interference signal, including: Based on the decoupled location and content query pair, the second fused feature is decoded to output an initial confidence score; The confidence score is determined based on the initial confidence score and the IOU confidence score predicted based on IOU.

[0105] For example, in an embodiment of this application, the identification unit 1103 is used to determine the confidence score based on the initial confidence score and the IOU confidence score predicted based on IOU, including: Determine the product of the initial credibility score and the IOU credibility score; The product is determined as the confidence score.

[0106] The UAV signal interference identification device 110 provided in this application embodiment can execute the technical solution of the UAV signal interference identification method in any of the above embodiments. Its implementation principle and beneficial effects are similar to those of the UAV signal interference identification method. Please refer to the implementation principle and beneficial effects of the UAV signal interference identification method. It will not be repeated here.

[0107] Figure 12 This is a schematic diagram of the physical structure of an electronic device provided in an embodiment of this application, such as... Figure 12 As shown, the electronic device may include: a processor 1210, a communications interface 1220, a memory 1230, and a communication bus 1240, wherein the processor 1210, the communications interface 1220, and the memory 1230 communicate with each other through the communication bus 1240. The processor 1210 can call logic instructions in the memory 1230 to execute a method for identifying interference in UAV signals. This method includes: acquiring the interfered UAV signal; performing a short-time Fourier transform on the UAV signal to obtain a two-dimensional time-frequency map corresponding to the UAV signal; inputting the two-dimensional time-frequency map into an interference category identification model, and outputting the interference type to which the interference signal belongs through the interference category identification model; and inputting the two-dimensional time-frequency map into an interference target detection model based on the parallel fusion of convolution operation paths and self-attention operation paths, and outputting a prediction box and a confidence score corresponding to the interference signal through the interference target detection model. The identification result of the interference signal includes the interference type, the prediction box, and the confidence score. The prediction box indicates the spatiotemporal position of the interference signal in the two-dimensional time-frequency map, and the confidence score indicates the confidence level of the prediction box.

[0108] Furthermore, the logical instructions in the aforementioned memory 1230 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0109] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the interference identification method for UAV signals provided by the above methods. The method includes: acquiring the interfered UAV signal; performing short-time Fourier transform on the UAV signal to obtain a two-dimensional time-frequency map corresponding to the UAV signal; inputting the two-dimensional time-frequency map into an interference category identification model, and outputting the interference type to which the interference signal belongs through the interference category identification model; and inputting the two-dimensional time-frequency map into an interference target detection model based on the parallel fusion of convolution operation path and self-attention operation path, and outputting the prediction box and confidence score corresponding to the interference signal through the interference target detection model. The identification result of the interference signal includes the interference type, the prediction box, and the confidence score. The prediction box is used to indicate the spatiotemporal position of the interference signal in the two-dimensional time-frequency map, and the confidence score is used to indicate the confidence of the prediction box.

[0110] In another aspect, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements an interference identification method for UAV signals provided by the methods described above. The method includes: acquiring an interfered UAV signal; performing a short-time Fourier transform on the UAV signal to obtain a two-dimensional time-frequency map corresponding to the UAV signal; inputting the two-dimensional time-frequency map into an interference category identification model, and outputting the interference type to which the interference signal belongs through the interference category identification model; and inputting the two-dimensional time-frequency map into an interference target detection model based on the parallel fusion of convolution operation path and self-attention operation path, and outputting a prediction box and a confidence score corresponding to the interference signal through the interference target detection model. The identification result of the interference signal includes the interference type, the prediction box, and the confidence score. The prediction box indicates the spatiotemporal position of the interference signal in the two-dimensional time-frequency map, and the confidence score indicates the confidence of the prediction box.

[0111] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0112] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for identifying interference in UAV signals, characterized in that, include: Acquire jammed drone signals; The UAV signal is converted and processed using short-time Fourier transform to obtain a two-dimensional time-frequency diagram corresponding to the UAV signal; The two-dimensional time-frequency image is input into the interference category identification model, which outputs the interference type to which the interference signal belongs; and the two-dimensional time-frequency image is input into the interference target detection model based on the parallel fusion of convolution operation path and self-attention operation path, which outputs the prediction box and confidence score corresponding to the interference signal. The identification result of the interference signal includes the interference type, the prediction box, and the confidence score. The prediction box is used to indicate the spatiotemporal position of the interference signal in the two-dimensional time-frequency diagram, and the confidence score is used to indicate the confidence of the prediction box.

2. The method according to claim 1, characterized in that, The process of converting the UAV signal using short-time Fourier transform to obtain a two-dimensional time-frequency graph corresponding to the UAV signal includes: A time-localized time window function is applied to the UAV signal, and a Fourier transform is performed on the signal segment within the current time window to obtain the spectral information of the signal segment corresponding to the current time window. The current time window is slid along the time axis to obtain the spectrum diagrams of the signal segments corresponding to each of the multiple time windows; wherein, the spectrum diagrams of the signal segments corresponding to each of the multiple time windows constitute the two-dimensional time-frequency diagram, and there is overlap between adjacent time windows.

3. The method according to claim 1 or 2, characterized in that, The interference target detection model includes a backbone network and a feature processing module; The step of inputting the two-dimensional time-frequency image into an interference target detection model based on the parallel fusion of convolution operation paths and self-attention operation paths, and outputting the prediction box and confidence score corresponding to the interference signal through the interference target detection model, includes: The two-dimensional time-frequency map is input into the backbone network based on the parallel fusion of convolution operation path and self-attention operation path. The backbone network performs multi-scale convolution on the two-dimensional time-frequency map and outputs the target multi-scale feature map corresponding to the two-dimensional time-frequency map. The target multi-scale features are input into the feature processing module, which performs feature interaction and fusion on the target multi-scale features and outputs the prediction box and confidence score corresponding to the interference signal.

4. The method according to claim 3, characterized in that, The backbone network includes a first high-dimensional feature extraction module and a second high-dimensional feature extraction module based on the parallel fusion of convolution operation path and self-attention operation path; The step of inputting the two-dimensional time-frequency map into the backbone network based on the parallel fusion of convolutional operation paths and self-attention operation paths, and performing multi-scale convolution on the two-dimensional time-frequency map through the backbone network to output the target multi-scale feature map corresponding to the two-dimensional time-frequency map includes: The two-dimensional time-frequency graph is input into the first high-dimensional feature extraction module, and the high-dimensional feature extraction module performs high-dimensional feature extraction on the two-dimensional time-frequency graph and outputs high-dimensional features. The high-dimensional features are input into the second high-dimensional feature extraction module, which performs parallel feature extraction based on convolution operation path and self-attention operation path, and fuses the parallel extracted features to output the target multi-scale feature map.

5. The method according to claim 4, characterized in that, The second high-dimensional feature extraction module includes multiple dynamically convolutional and self-attention bottleneck modules connected in series. The high-dimensional features are input into the second high-dimensional feature extraction module, which is based on the parallel fusion of convolutional operation paths and self-attention operation paths. The second high-dimensional feature extraction module performs parallel feature extraction on the high-dimensional features based on convolutional operation paths and self-attention operation paths, and fuses the parallel extracted features to output the target multi-scale feature map, including: The high-dimensional features are input into the first dynamic convolution and self-attention bottleneck module. The high-dimensional features are extracted in parallel based on the convolution operation path and the self-attention operation path through the first dynamic convolution and self-attention bottleneck module. The extracted features are then fused to output the first multi-scale feature map. For the other dynamic convolution and self-attention bottleneck modules besides the first one, parallel feature extraction based on convolution operation path and self-attention operation path is performed on the second multi-scale feature map output by the previous dynamic convolution and self-attention bottleneck module through the other dynamic convolution and self-attention bottleneck modules, and the parallel extracted features are fused to output a third multi-scale feature map; wherein, the third multi-scale feature map output by the last dynamic convolution and self-attention bottleneck module among the other dynamic convolution and self-attention bottleneck modules is the target multi-scale feature map.

6. The method according to claim 5, characterized in that, The first dynamic convolution and self-attention bottleneck module includes a convolution projection unit, a convolution unit based on the convolution operation path, and a self-attention unit based on the self-attention operation path. The high-dimensional features are input into the first dynamic convolution and self-attention bottleneck module. This module performs parallel feature extraction on the high-dimensional features based on convolution and self-attention operation paths. The extracted features are then fused to output a first multi-scale feature map, including: The high-dimensional features are input into the convolutional projection unit, and the high-dimensional features are linearly projected through the convolutional projection unit to output the query, key, and value. The query, key, and value are respectively input into the convolutional unit and the self-attention unit. The convolutional unit performs a depthwise separable convolution operation on the query, key, and value, and outputs a convolutional feature map. The self-attention unit performs a multi-head self-attention operation on the query, key, and value, and outputs a self-attention feature map. The convolutional feature map and the self-attention feature map are fused to obtain the first multi-scale feature map.

7. The method according to claim 6, characterized in that, The process of fusing the convolutional feature map and the self-attention feature map to obtain the first multi-scale feature map includes: Determine the gating coefficients corresponding to the convolutional unit and the self-attention unit respectively; Based on the gating coefficients corresponding to the convolutional unit and the self-attention unit, the convolutional feature map and the self-attention feature map are weighted and fused to obtain the first multi-scale feature map.

8. The method according to claim 3, characterized in that, The feature processing module includes an attention-based intra-scale feature interaction (AIFI) module, a cross-scale feature fusion (CFFM) module, and a decoding module. The step of inputting the multi-scale features of the target into the feature processing module, performing feature interaction and fusion on the multi-scale features of the target through the feature processing module, and outputting the prediction box and confidence score corresponding to the interference signal includes: The target multi-scale feature map is input into the AIFI module, and the AIFI module performs intra-scale feature self-attention fusion on the target multi-scale feature map to output the first fused feature. Both the target multi-scale feature map and the first fused feature are input into the CFFM. The CFFM is used to fuse the target multi-scale feature map and the first fused feature to obtain the second fused feature. The second fused feature is input into the decoding module, which decodes the second fused feature based on decoupled location and content queries, and outputs the prediction box and confidence score corresponding to the interference signal.

9. The method according to claim 8, characterized in that, The step of decoding the second fused feature based on the decoupled location and content query by the decoding module and outputting the confidence score corresponding to the interference signal includes: Based on the decoupled location and content query pair, the second fused feature is decoded to output an initial confidence score; The confidence score is determined based on the initial confidence score and the IOU confidence score predicted based on IOU.

10. The method according to claim 9, characterized in that, The process of determining the credibility score based on the initial credibility score and the IOU-predicted credibility score includes: Determine the product of the initial credibility score and the IOU credibility score; The product is determined as the confidence score.

Images