Weak unmanned aerial vehicle flight control signal detection method and system
By generating gradient and corner feature maps and combining them with multi-channel input tensors, a target detection network was used to solve the accuracy problem of UAV flight control signal detection in complex electromagnetic environments, achieving efficient signal recognition and localization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUAXIN CONSULTATING CO LTD
- Filing Date
- 2026-04-24
- Publication Date
- 2026-07-31
AI Technical Summary
In complex electromagnetic environments, existing methods struggle to accurately detect UAV flight control signals under extremely low signal-to-noise ratio conditions, resulting in high false alarm rates, high false alarm rates, and inaccurate positioning.
The original time-frequency map is generated by short-time Fourier transform, and gradient feature maps and corner feature maps are obtained by convolution calculation. Combined with multi-channel input tensors, an object detection network is used for localization and recognition.
It improves the detection rate of extremely short signals, reduces the false alarm and missed detection rates, ensures the accuracy of signal positioning, and adapts to the detection of UAV flight control signals in complex electromagnetic environments.
Smart Images
Figure CN122119804B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of signal detection, and in particular to methods and systems for detecting weak UAV flight control signals. Background Technology
[0002] In complex electromagnetic environments, there are two main methods for detecting and locating UAV flight control signals. One method targets broadband image transmission signals, utilizing their large bandwidth and long duration characteristics to achieve identification by detecting cyclic prefixes or specific pilot sequences. However, this method cannot handle UAVs in silent flight mode with image transmission turned off. The other method targets narrowband flight control signals, employing frequency hopping technology, which appears as discrete short spots on the time-frequency graph. Existing methods often directly use general target detection networks (such as Faster R-CNN and YOLO) to treat these signals as small targets for detection.
[0003] However, in the complex electromagnetic environment of cities, due to energy flooding, the flight control signal energy may be lower than the peak background noise under extremely low signal-to-noise ratio conditions (e.g., -15dB), making it almost invisible on the time-frequency plot. General detection networks rely on grayscale values (signal energy) to extract features, and their convolutional backbone networks struggle to extract effective signal contours from noise, leading to a high false alarm rate. Secondly, noise texture interference is severe. The random speckle textures generated by Gaussian white noise or Rayleigh fading channels on the time-frequency plot are highly similar in morphology to frequency-hopping signals. Traditional convolutional networks struggle to distinguish between regular signal speckles and random noise speckles, resulting in a significant increase in the false alarm rate. Furthermore, multipath effects cause signal edge blurring; existing methods are inaccurate in boundary localization when regressing the signal center frequency and bandwidth, directly affecting the accuracy of subsequent interference countermeasures. Summary of the Invention
[0004] This application provides a method, system, electronic device, and storage medium for detecting weak UAV flight control signals, in order to at least solve the problem of inaccurate detection of UAV flight control signals in low signal-to-noise ratio environments in related technologies.
[0005] Firstly, embodiments of this application provide a method for detecting weak UAV flight control signals, the method comprising:
[0006] Acquire radio frequency I / Q data, perform short-time Fourier transform on the radio frequency I / Q data to generate an original time-frequency map, perform convolution calculation on the original time-frequency map to obtain its gradient components in the time axis direction and gradient components in the frequency axis direction, and generate a gradient feature map.
[0007] Based on the gradient components, the structure tensor matrix of each pixel in the original time-frequency image is determined, and the corner response value of each pixel is calculated using the structure tensor matrix to generate a corner feature map.
[0008] The original time-frequency map, the gradient feature map, and the corner feature map are concatenated to form a multi-channel input tensor;
[0009] Based on the target detection network, the UAV flight control frequency hopping signal instances in the original time-frequency graph are located and identified according to the multi-channel input tensor.
[0010] In some embodiments, the step of performing convolution calculations on the original time-frequency map to obtain its gradient components along the time axis and frequency axis, respectively, to generate a gradient feature map, includes:
[0011] A two-dimensional differential gradient operator is used to perform a convolution operation on the original time-frequency graph to calculate the time gradient graph in the time axis direction and the frequency gradient graph in the frequency axis direction of the original time-frequency graph, respectively.
[0012] Based on the time gradient map and the frequency gradient map, determine the gradient magnitude map and the gradient direction map;
[0013] The gradient magnitude map is subjected to nonmaximum suppression processing based on the gradient direction map to obtain the gradient feature map, wherein the two-dimensional differential gradient operator is a Sobel operator or a Scharr operator, and the time axis gradient operator and the frequency axis gradient operator in the two-dimensional differential gradient operator are orthogonal to each other.
[0014] In some embodiments, the step of calculating the corner response values of each pixel using the structure tensor matrix to generate a corner feature map includes:
[0015] For each pixel in the original time-frequency graph, take its neighborhood window and calculate the structure tensor matrix;
[0016] A corner response function is constructed based on the Harris corner criterion, and the corner response value of each pixel is calculated based on the corner response function.
[0017] Thresholding and local maxima suppression are applied to the corner response values to generate a binarized corner feature map.
[0018] In some embodiments, the corner response function includes:
[0019] R=det(M)-k*(trace(M)) 2
[0020] Where R is the corner response value; M is the structure tensor matrix; k is an empirical constant; det(M) is the determinant of M; and trace(M) is the trace of M, which is the sum of the main diagonal elements of M.
[0021] In some embodiments, performing a short-time Fourier transform on the RF I / Q data to generate the original time-frequency diagram includes:
[0022] A sliding window strategy with a window overlap rate greater than 75% is used to perform a short-time Fourier transform on the RF I / Q data to generate the original time-frequency diagram;
[0023] Based on logarithmic operations, the amplitude values of the original time-frequency diagram are dynamically compressed.
[0024] In some embodiments, the target detection network is a small target detection network based on the Anchor-Free mechanism, with its backbone network adopting a CSPDarknet structure with cross-stage partial connections and its neck network adopting a path aggregation network structure.
[0025] In some embodiments, the step of locating and identifying instances of UAV flight control frequency hopping signals in the original time-frequency graph based on the target detection network and the multi-channel input tensor includes:
[0026] The multi-channel input tensor is input into the target detection network, and the output is a detection result containing the location information of the signal time-frequency boundary and the category confidence. The loss function used for training the target detection network is the Focal Loss loss function.
[0027] Secondly, embodiments of this application provide a weak UAV flight control signal detection system, the system comprising:
[0028] The first feature map generation module is used to collect radio frequency I / Q data, perform short-time Fourier transform on the radio frequency I / Q data to generate an original time-frequency map, perform convolution calculation on the original time-frequency map to obtain its gradient component in the time axis direction and its gradient component in the frequency axis direction, and generate a gradient feature map.
[0029] The second feature map generation module is used to determine the structure tensor matrix of each pixel in the original time-frequency map based on the gradient components, calculate the corner response value of each pixel through the structure tensor matrix, and generate a corner feature map.
[0030] The splicing module is used to splice the original time-frequency map, the gradient feature map, and the corner feature map to form a multi-channel input tensor;
[0031] The detection module is used to locate and identify instances of UAV flight control frequency hopping signals in the original time-frequency graph based on the target detection network and the multi-channel input tensor.
[0032] Thirdly, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the weak UAV flight control signal detection method described in the first aspect above.
[0033] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the weak UAV flight control signal detection method as described in the first aspect above.
[0034] Compared to related technologies, the weak UAV flight control signal detection method provided in this application suppresses the low-frequency background substrate and highlights the signal outline through gradient feature maps; the corner feature map transforms the frequency hopping spots that originally occupied a certain area into 4 bright pixels, thereby improving the detection rate of extremely short signals; and the introduction of explicit frequency axis gradients gives the network a clear reference boundary when regressing the signal bandwidth, making the output positioning results more consistent with the real signal.
[0035] Multi-dimensional feature fusion (original time-frequency map + gradient feature map + corner feature map) enables the network to accurately identify the structural features of flight control signals from noisy backgrounds, avoiding problems such as missed detections, false alarms and inaccurate positioning in complex scenarios with low signal-to-noise ratio and high noise interference. Attached Figure Description
[0036] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:
[0037] Figure 1 This is a flowchart of a weak UAV flight control signal detection method according to an embodiment of this application;
[0038] Figure 2 This is a flowchart of a gradient-enhanced detection method according to an embodiment of this application;
[0039] Figure 3 This is a schematic diagram illustrating the gradient feature extraction principle according to an embodiment of this application;
[0040] Figure 4 This is a diagram illustrating the effect of structural tensor corner detection according to an embodiment of this application.
[0041] Figure 5 This is a diagram of an improved SMNet detection network architecture according to an embodiment of this application;
[0042] Figure 6This is a comparison chart of detection results under different preprocessing methods according to embodiments of this application;
[0043] Figure 7 This is an example diagram showing the detection results of an algorithm according to an embodiment of this application in a real-world unauthorized flight scenario;
[0044] Figure 8 This is a structural block diagram of a weak UAV flight control signal detection system according to an embodiment of this application;
[0045] Figure 9 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application. Detailed Implementation
[0046] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.
[0047] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.
[0048] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.
[0049] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.
[0050] This embodiment provides a method for detecting weak UAV flight control signals. Figure 1 This is a flowchart of a weak UAV flight control signal detection method according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps:
[0051] Step S101: Acquire radio frequency I / Q data, perform short-time Fourier transform on the radio frequency I / Q data to generate the original time-frequency map, perform convolution calculation on the original time-frequency map to obtain its gradient components in the time axis direction and the gradient components in the frequency axis direction, and generate a gradient feature map.
[0052] In some embodiments, step S101, which involves performing a short-time Fourier transform on the RF I / Q data to generate the original time-frequency diagram, includes:
[0053] Step S1011: Using a sliding window strategy with a window overlap rate greater than 75%, perform short-time Fourier transform on the RF I / Q data to generate the original time-frequency diagram.
[0054] Step S1012: Based on logarithmic operations, the amplitude values of the original time-frequency diagram are dynamically compressed.
[0055] RF I / Q data is acquired and STFT transformation is performed. To capture microsecond-level frequency hopping signal edges, a sliding window strategy with high overlap (Overlap > 75%) is adopted to generate a fine-grained time-frequency map with extremely high time resolution.
[0056] The high-overlap sliding window strategy with a window overlap rate of more than 75% reduces the time interval between adjacent sliding windows, enabling precise capture of the edge features of microsecond-level frequency hopping signals, ensuring that the time boundary of the signal is completely recorded, and significantly improving the time resolution of the time-frequency map; each time-frequency point overlaps with multiple sliding windows before and after it, significantly enhancing the continuity and sharpness of the signal edges, which helps to identify the presence of signals under low signal-to-noise ratio.
[0057] Logarithmic dynamic range compression can compress the amplitude values of the time-frequency plot to a reasonable range, making it easier for the network to learn signal features. Simultaneously, the improved temporal resolution of the high-overlap-rate time-frequency plot makes signal edge features more prominent, and the compressed time-frequency plot can more accurately reflect the true form of the signal.
[0058] In some embodiments, step S101 involves performing convolution calculations on the original time-frequency map to obtain its gradient components along the time axis and frequency axis, respectively, thereby generating a gradient feature map.
[0059] Step S1013: The original time-frequency image is convolved using a two-dimensional differential gradient operator to calculate the time gradient image of the original time-frequency image in the time axis direction and the frequency gradient image in the frequency axis direction, respectively.
[0060] Step S1014: Based on the time gradient map and the frequency gradient map, determine the gradient magnitude map and the gradient direction map;
[0061] Step S1015: Perform nonmaximum suppression processing on the gradient magnitude map based on the gradient direction map to obtain the gradient feature map. The two-dimensional differential gradient operator is either the Sobel operator or the Scharr operator. The time axis gradient operator and the frequency axis gradient operator in the two-dimensional differential gradient operator are orthogonal to each other.
[0062] For the input time-frequency graph S raw Calculate the time axis gradient G respectively t and frequency axis gradient G f .
[0063] Convolution calculations are performed using the Scharr or Sobel operators, which have better noise resistance. , , where K t and K f This is the gradient operator. The high response region of Gt corresponds to the start and end times (energy abrupt change) of the signal; G fThe high response region corresponds to the upper and lower sidebands (bandwidth boundaries) of the signal.
[0064] Using mutually orthogonal time axis gradient operators K t and frequency axis gradient operator K f This achieves two-dimensional orthogonal decomposition of the time-frequency graph. K t Precisely capturing signal changes along the time axis, its high-response region directly corresponds to the start / end moments of the signal (energy mutation points), providing crucial information for locating the time boundaries of frequency-hopping signals. f It accurately captures signal variations along the frequency axis, with its high-response region corresponding to the upper / lower sidebands (bandwidth boundaries) of the signal, providing support for accurate measurement of signal bandwidth. The orthogonal design ensures decoupling of features in the time and frequency dimensions, avoiding cross-dimensional interference and significantly improving the specificity of gradient features.
[0065] Step S102: Determine the structure tensor matrix of each pixel in the original time-frequency map based on the gradient components, calculate the corner response value of each pixel through the structure tensor matrix, and generate a corner feature map.
[0066] In some embodiments, step S102 specifically includes:
[0067] Step S1021: For each pixel in the original time-frequency map, take its neighborhood window and calculate the structure tensor matrix;
[0068] Step S1022: Construct a corner response function based on the Harris corner criterion, and calculate the corner response value of each pixel based on the corner response function;
[0069] Step S1023: Threshold the corner response values and suppress local maxima to generate a binarized corner feature map.
[0070] To further lock onto the frequency-hopping signal (usually in the form of a rectangular block), the structure tensor matrix M for each pixel is calculated:
[0071]
[0072] Where w is the Gaussian weighted window.
[0073] In some embodiments, the corner response function includes:
[0074] R=det(M)-k*(trace(M)) 2
[0075] Where R is the corner response value; M is the structure tensor matrix; k is an empirical constant; det(M) is the determinant of M; and trace(M) is the trace of M, which is the sum of the main diagonal elements of M.
[0076] The generated corner feature map S corner It has an extremely high response only at the four vertices (time-frequency corners) of the frequency-hopping signal, and almost no response to stationary noise.
[0077] The Harris corner response function is used to effectively suppress noise-induced fluctuations in corner response values by calculating the difference between the determinant and the trace. The selection of the empirical constant k can be adjusted according to the actual scenario, improving the robustness of corner detection. The Harris corner criterion ensures the stability of corner response values by weighting the determinant and trace of the structural tensor matrix, providing a reliable basis for subsequent thresholding and local maximum suppression.
[0078] The rectangular block structure of a frequency-hopping signal generates four distinct high-response points in the corner feature map. These points accurately reflect the signal's boundary and shape. Binarization of the corner feature map enables the network to quickly identify the signal's presence, significantly improving detection efficiency and robustness. By combining the structural tensor matrix with the Harris corner criterion, precise localization of the rectangular block structure of the frequency-hopping signal is achieved.
[0079] Step S103: The original time-frequency map, gradient feature map and corner feature map are spliced together to form a multi-channel input tensor.
[0080] Original time-frequency diagram Gradient feature map and corner feature map The input tensor is concatenated along the channel dimension to form a 3-channel input tensor. Multi-channel fusion of corner feature maps, gradient feature maps, and the original time-frequency map further enhances the network's ability to recognize frequency-hopping signals.
[0081] By concatenating the original time-frequency map (energy dimension), gradient feature map (structure / edge dimension), and corner feature map (critical corner dimension) into a 3-channel tensor, complementary multi-dimensional features are achieved: the original time-frequency map provides the basic energy distribution of the signal, ensuring that the network does not miss potential signal regions; the gradient feature map strengthens the time-frequency boundaries of the signal (time start and end / frequency bandwidth), solving the edge blurring problem caused by multipath effects; and the corner feature map locks the rectangular corners of the signal, thoroughly distinguishing regular signal spots from random noise spots. After the three are fused, the network can sense both energy and recognize structure, significantly reducing the false alarm / missed detection rate.
[0082] Step S104: Based on the target detection network, locate and identify the UAV flight control frequency hopping signal instances in the original time-frequency map according to the multi-channel input tensor.
[0083] The input is fed into an improved object detection network (such as SMNet). Optionally, the first layer of the network uses a large convolutional kernel (7×7) to quickly fuse gradient and energy information.
[0084] The first layer of the network uses a 7×7 large convolutional kernel to cover a larger time-frequency region at once, quickly capturing the rectangular block-like global structure of the frequency hopping signal. Energy information and gradient / corner structural information can be deeply fused in the first layer, avoiding information fragmentation caused by feature layering. The frequency hopping signal is a short spot (small target) at the microsecond level, and the large convolutional kernel can reduce the number of downsampling times and retain more detailed features.
[0085] In some embodiments, the target detection network is a small target detection network based on the Anchor-Free mechanism, with its backbone network adopting a CSPDarknet structure with cross-stage partial connections and its neck network adopting a path aggregation network structure.
[0086] The Anchor-Free mechanism eliminates the need for pre-defined anchors, directly regressing the time-frequency boundaries of the signal. This adapts to the discrete, short, and variable-size characteristics of frequency-hopping signals, avoiding missed detections caused by anchor mismatch. The CSPDarknet backbone reduces computational load in cross-stage partial connection structures while enhancing feature reuse, preventing the loss of small target features in deep networks. The Path Aggregation Network (PAN) neck integrates top-down and bottom-up feature fusion, combining shallow detail features (corners / gradients) with deep semantic features (energy) to improve the accuracy of small target detection.
[0087] In some embodiments, step S104 includes: inputting a multi-channel input tensor into an object detection network and outputting a detection result containing location information of the signal time-frequency boundary and class confidence, wherein the loss function used for training the object detection network is the Focal Loss loss function.
[0088] The introduction of IoU-aware Loss focuses on penalizing inaccurate boundary regression, forcing the network to accurately lock the signal edges using gradient information.
[0089] The Focal Loss function addresses the imbalance between positive and negative samples by assigning higher weights to difficult-to-distinguish samples (low signal-to-noise ratio signals) and lower weights to easily distinguishable samples (noise), thereby improving the detection rate of weak signals. IoU-aware Loss directly uses the IoU between the predicted bounding box and the ground truth bounding box as the optimization objective, focusing on penalizing inaccurate boundary regression. This forces the network to utilize edge information in the gradient feature map to accurately lock the time-frequency boundaries (start and end times / upper and lower sidebands) of the signal, ensuring the accuracy of subsequent interference countermeasures.
[0090] Figure 2 This is a flowchart of a gradient-enhanced detection method according to an embodiment of this application.
[0091] Through the above steps, the gradient feature map suppresses the low-frequency background base and highlights the signal outline; the corner feature map transforms the frequency-hopping spots that originally occupied a certain area into 4 bright pixels, thereby improving the detection rate of extremely short signals; and by introducing an explicit frequency axis gradient, the network has a clear reference boundary when regressing the signal bandwidth, making the output localization results more consistent with the real signal.
[0092] Multi-dimensional feature fusion (original time-frequency map + gradient feature map + corner feature map) enables the network to accurately identify the structural features of flight control signals from noisy backgrounds, avoiding problems such as missed detections, false alarms and inaccurate positioning in complex scenarios with low signal-to-noise ratio and high noise interference.
[0093] The detection process implementation example is as follows:
[0094] Step S201, signal acquisition and preprocessing.
[0095] An AD9361 RF front-end was used, with a center frequency of 2.44 GHz and a bandwidth of 40 MHz; a sampling rate of 50 MSps, an STFT window length of N=2048, and a Hop Size of 256 (overlap rate of 87.5%) to acquire RF signals. A high-resolution 512×512 time-frequency grayscale image was generated based on the acquired data. At this point, the signal appears very blurry.
[0096] Step S202, Gradient Feature Engineering (Core Step).
[0097] The Scharr operator is used because it is more rotationally symmetric and more sensitive to small gradients than Sobel.
[0098] Convolution is performed on the normalized time-frequency graph S to obtain G. x G y .
[0099] Calculate the gradient magnitude In the feature map obtained in this step, the signal outline will show a double-sided delineation effect.
[0100] Figure 3 This is a schematic diagram illustrating the gradient feature extraction principle according to an embodiment of this application; Figure 4 This is a diagram illustrating the effect of structural tensor corner detection according to an embodiment of this application.
[0101] Step S203, Neural Network Construction.
[0102] The lightweight CSPDarknet (derived from YOLOv5 / 7) is used to ensure real-time performance.
[0103] The input layer was modified by changing the number of input channels of the first convolutional kernel of the network from 1 to 3 (corresponding to the original energy channel, gradient magnitude channel, and gradient direction channel respectively).
[0104] PANet (Path Aggregation Network) is used for multi-scale feature fusion to ensure that both extremely short frequency-hopping signals (small scale) and longer burst signals (large scale) can be detected.
[0105] The output layer predicts three predefined reference boxes, each corresponding to a frequency hopping signal shape with a different aspect ratio.
[0106] Figure 5 This is a diagram of an improved SMNet detection network architecture according to an embodiment of this application.
[0107] Step S204, Training and Reasoning
[0108] Manually label the time-frequency bounding boxes (tstart, flow, tend, fhigh) of the real signal; use Mosaic data augmentation to prevent overfitting; output detection boxes with confidence > 0.5.
[0109] Example of an FPGA-based real-time edge detection acceleration method:
[0110] For scenarios requiring real-time generation of gradient maps via hardware logic, a line buffer is allocated within the FPGA to cache three lines of STFT spectrum data; a 3×3 register array is used to implement the sliding window convolution of the Scharr operator; the gradient module performs differential operations immediately while the STFT module outputs data, with a delay of only 3 clock cycles.
[0111] A gradient threshold Thgrad is set in the hardware. Only when the gradient magnitude is greater than Thgrad is the pixel marked as a "Region of Interest (ROI)". Subsequent recognition algorithms only process the ROI region, which greatly reduces the amount of data processing.
[0112] Table 1 is an example table of key parameters involved in an algorithm according to an embodiment of this application.
[0113] Table 1
[0114] Figure 6 This is a comparison chart of detection results under different preprocessing methods according to embodiments of this application; Figure 7 This is an example diagram showing the detection results of an algorithm according to an embodiment of this application in a real-world unauthorized flight scenario.
[0115] Figure 7 This paper demonstrates the detection performance of this invention in a real, complex electromagnetic environment. The bright vertical broadband stripes in the figure represent strong Wi-Fi burst interference (SIR < -10dB) present in the environment, while the scattered weak dark spots (highlighted in green) represent the flight control frequency-hopping signals of the target UAV. Thanks to the gradient feature enhancement mechanism of this invention, although the absolute energy of the Wi-Fi signal is much higher than that of the UAV signal, its time-frequency gradient characteristics differ significantly from those of the frequency-hopping signal (Wi-Fi has no obvious upper or lower frequency domain edges). Therefore, the detection network can successfully suppress high-energy interference, accurately locate and select all weak flight control signal pulses, achieving a robust identification effect that detects only the target and not the interference.
[0116] It should be noted that the steps shown in the above process or in the flowchart of the accompanying figures can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0117] This embodiment also provides a weak UAV flight control signal detection system, which is used to implement the above embodiments and preferred embodiments, and will not be repeated as already described. As used below, the terms "module," "unit," "subunit," etc., can refer to a combination of software and / or hardware that performs a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.
[0118] Figure 8 This is a structural block diagram of a weak UAV flight control signal detection system according to an embodiment of this application, as shown below. Figure 8 As shown, the system includes:
[0119] The first feature map generation module 81 is used to acquire radio frequency I / Q data, perform short-time Fourier transform on the radio frequency I / Q data to generate the original time-frequency map, perform convolution calculation on the original time-frequency map to obtain its gradient component in the time axis direction and its gradient component in the frequency axis direction, and generate a gradient feature map.
[0120] The second feature map generation module 82 is used to determine the structure tensor matrix of each pixel in the original time-frequency map based on the gradient components, calculate the corner response value of each pixel through the structure tensor matrix, and generate a corner feature map.
[0121] The splicing module 83 is used to splice the original time-frequency map, gradient feature map and corner feature map to form a multi-channel input tensor.
[0122] The detection module 84 is used to locate and identify instances of UAV flight control frequency hopping signals in the original time-frequency graph based on the target detection network and the multi-channel input tensor.
[0123] In some embodiments, the first feature map generation module 81 includes:
[0124] The frequency gradient map generation module is used to perform a convolution operation on the original time-frequency map using a two-dimensional differential gradient operator, and to calculate the time gradient map of the original time-frequency map in the time axis direction and the frequency gradient map in the frequency axis direction, respectively.
[0125] The gradient graph processing module is used to determine the gradient magnitude graph and gradient direction graph based on the time gradient graph and the frequency gradient graph.
[0126] The gradient feature map generation module is used to perform nonmaximum suppression processing on the gradient magnitude map based on the gradient direction map to obtain the gradient feature map. The two-dimensional differential gradient operator is either the Sobel operator or the Scharr operator. The time axis gradient operator and the frequency axis gradient operator in the two-dimensional differential gradient operator are orthogonal to each other.
[0127] In some embodiments, the second feature map generation module 82 includes:
[0128] The matrix construction module is used to calculate the structure tensor matrix by taking the neighborhood window of each pixel in the original time-frequency graph.
[0129] The corner calculation module is used to construct corner response functions based on the Harris corner criterion, and to calculate the corner response value of each pixel based on the corner response functions.
[0130] The corner feature map generation module is used to threshold and suppress local maxima of corner response values to generate binarized corner feature maps.
[0131] In some embodiments, the corner response function includes:
[0132] R=det(M)-k*(trace(M)) 2
[0133] Where R is the corner response value; M is the structure tensor matrix; k is an empirical constant; det(M) is the determinant of M; and trace(M) is the trace of M, which is the sum of the main diagonal elements of M.
[0134] In some embodiments, the first feature map generation module 81 includes:
[0135] The original time-frequency diagram generation module is used to perform short-time Fourier transform on RF I / Q data using a sliding window strategy with a window overlap rate greater than 75% to generate the original time-frequency diagram.
[0136] The compression module is used to dynamically compress the amplitude values of the original time-frequency graph based on logarithmic operations.
[0137] In some embodiments, the target detection network is a small target detection network based on the Anchor-Free mechanism, with its backbone network adopting a CSPDarknet structure with cross-stage partial connections and its neck network adopting a path aggregation network structure.
[0138] In some embodiments, the detection module 84 includes: an identification module, used to input a multi-channel input tensor into the target detection network and output a detection result containing the location information of the signal time-frequency boundary and the category confidence, wherein the loss function used for training the target detection network is the Focal Loss loss function.
[0139] Through the above system, the gradient feature map suppresses the low-frequency background base and highlights the signal outline; the corner feature map transforms the frequency-hopping spots that originally occupied a certain area into 4 bright pixels, thereby improving the detection rate of extremely short signals; and the introduction of an explicit frequency axis gradient gives the network a clear reference boundary when regressing the signal bandwidth, making the output localization results more in line with the real signal.
[0140] Multi-dimensional feature fusion (original time-frequency map + gradient feature map + corner feature map) enables the network to accurately identify the structural features of flight control signals from noisy backgrounds, avoiding problems such as missed detections, false alarms and inaccurate positioning in complex scenarios with low signal-to-noise ratio and high noise interference.
[0141] It should be noted that the above modules can be functional modules or program modules, and can be implemented through software or hardware. For modules implemented through hardware, the above modules can reside in the same processor; or the above modules can be located in different processors in any combination.
[0142] This embodiment also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.
[0143] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.
[0144] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:
[0145] S1: Acquire RF I / Q data, perform short-time Fourier transform on the RF I / Q data to generate the original time-frequency map, perform convolution calculation on the original time-frequency map to obtain its gradient components in the time axis direction and the gradient components in the frequency axis direction, and generate a gradient feature map.
[0146] S2, determine the structure tensor matrix of each pixel in the original time-frequency map based on the gradient components, calculate the corner response value of each pixel through the structure tensor matrix, and generate a corner feature map.
[0147] S3 concatenates the original time-frequency map, gradient feature map, and corner feature map to form a multi-channel input tensor.
[0148] S4, based on the target detection network, locates and identifies instances of UAV flight control frequency hopping signals in the original time-frequency graph according to the multi-channel input tensor.
[0149] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.
[0150] In one embodiment, Figure 9 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application, such as... Figure 9 As shown, an electronic device is provided, which can be a server, and its internal structure diagram can be as follows. Figure 9As shown, this electronic device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for detecting weak UAV flight control signals.
[0151] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.
[0152] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0153] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0154] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.
Claims
1. A weak unmanned aerial vehicle flight control signal detection method, characterized in that, The method includes: Acquire radio frequency I / Q data, perform short-time Fourier transform on the radio frequency I / Q data to generate an original time-frequency map, perform convolution calculation on the original time-frequency map to obtain its gradient components in the time axis direction and gradient components in the frequency axis direction, and generate a gradient feature map. Based on the gradient components, the structure tensor matrix of each pixel in the original time-frequency image is determined, and the corner response value of each pixel is calculated using the structure tensor matrix to generate a corner feature map. The original time-frequency map, the gradient feature map, and the corner feature map are concatenated to form a multi-channel input tensor; Based on the target detection network, the UAV flight control frequency hopping signal instances in the original time-frequency graph are located and identified according to the multi-channel input tensor.
2. The method of claim 1, wherein, The step of performing convolution calculation on the original time-frequency map to obtain its gradient components along the time axis and frequency axis, and generating a gradient feature map, includes: A two-dimensional differential gradient operator is used to perform a convolution operation on the original time-frequency graph to calculate the time gradient graph in the time axis direction and the frequency gradient graph in the frequency axis direction of the original time-frequency graph, respectively. Based on the time gradient map and the frequency gradient map, determine the gradient magnitude map and the gradient direction map; The gradient magnitude map is subjected to nonmaximum suppression processing based on the gradient direction map to obtain the gradient feature map, wherein the two-dimensional differential gradient operator is a Sobel operator or a Scharr operator, and the time axis gradient operator and the frequency axis gradient operator in the two-dimensional differential gradient operator are orthogonal to each other.
3. The method of claim 1, wherein, The step of calculating the corner response value of each pixel using the structure tensor matrix to generate a corner feature map includes: For each pixel in the original time-frequency graph, take its neighborhood window and calculate the structure tensor matrix; A corner response function is constructed based on the Harris corner criterion, and the corner response value of each pixel is calculated based on the corner response function. Thresholding and local maxima suppression are applied to the corner response values to generate a binarized corner feature map.
4. The method according to claim 3, characterized in that, The corner response function includes: R = det(M) - k * (trace(M)) 2 Where R is the corner response value; M is the structure tensor matrix; k is an empirical constant; det(M) is the determinant of M; and trace(M) is the trace of M, which is the sum of the main diagonal elements of M.
5. The method according to claim 1, characterized in that, The step of performing a short-time Fourier transform on the RF I / Q data to generate the original time-frequency diagram includes: A sliding window strategy with a window overlap rate greater than 75% is used to perform a short-time Fourier transform on the RF I / Q data to generate the original time-frequency diagram; Based on logarithmic operations, the amplitude values of the original time-frequency diagram are dynamically compressed.
6. The method according to claim 1, characterized in that, The target detection network is a small target detection network based on the Anchor-Free mechanism. Its backbone network adopts the CSPDarknet structure with cross-stage partial connections, and the neck network adopts the path aggregation network structure.
7. The method according to claim 1, characterized in that, The step of locating and identifying UAV flight control frequency hopping signal instances in the original time-frequency graph based on the target detection network and the multi-channel input tensor includes: The multi-channel input tensor is input into the target detection network, and the output is a detection result containing the location information of the signal time-frequency boundary and the category confidence. The loss function used for training the target detection network is the FocalLoss loss function.
8. A weak UAV flight control signal detection system, characterized in that, The system includes: The first feature map generation module is used to collect radio frequency I / Q data, perform short-time Fourier transform on the radio frequency I / Q data to generate an original time-frequency map, perform convolution calculation on the original time-frequency map to obtain its gradient component in the time axis direction and its gradient component in the frequency axis direction, and generate a gradient feature map. The second feature map generation module is used to determine the structure tensor matrix of each pixel in the original time-frequency map based on the gradient components, calculate the corner response value of each pixel through the structure tensor matrix, and generate a corner feature map. The splicing module is used to splice the original time-frequency map, the gradient feature map, and the corner feature map to form a multi-channel input tensor; The detection module is used to locate and identify instances of UAV flight control frequency hopping signals in the original time-frequency graph based on the target detection network and the multi-channel input tensor.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the weak UAV flight control signal detection method as described in any one of claims 1 to 7.
10. A storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the weak UAV flight control signal detection method as described in any one of claims 1 to 7.