Unmanned aerial vehicle spectrum anomaly detection and trajectory generation method based on unsupervised learning

Abstract

The application discloses a kind of unmanned aerial vehicle spectrum anomaly detection and trajectory generation method based on unsupervised learning, including by SDR continuous acquisition broadband radio frequency signal, obtain IQ data stream;Short-time Fourier transform is carried out, and time-frequency spectrogram is obtained;Robust normalization processing is carried out to time-frequency spectrogram, and after inhibiting slow change drift, standardization spectrogram is obtained;Constant false alarm rate detection is executed, and preliminary anomaly candidate binary mask is generated;From standardization spectrogram, the spectrogram sub-block corresponding to preliminary anomaly candidate binary mask is identified, input into pre-trained unsupervised deep learning model, and abnormal score chart is obtained;With preliminary anomaly candidate binary mask is fused, and generates abnormal heat map;Abnormal heat map is clustered in time-frequency dimension, and trajectory correlation and filtering are carried out to clustering cluster across time window, and the time-frequency trajectory of unmanned aerial vehicle signal is output.The application enhances the reliability of anomaly detection, reduces the risk of false detection and missed detection, and guarantees the continuity and accuracy of trajectory generation.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and in particular to a method for detecting UAV spectrum anomalies and generating trajectories based on unsupervised learning. Background Technology

[0002] In recent years, the widespread adoption of consumer and industrial-grade small drones has brought new security challenges. These drones often use Wi-Fi / ISM bands and amateur bands for communication, and widely employ technologies such as frequency hopping, spectrum broadening, and 4G / 5G cellular backhaul to circumvent traditional monitoring and interference. In existing anti-drone technology systems, radio frequency spectrum monitoring, as a passive and covert detection method, has become a key area of ​​development. Traditional spectrum monitoring methods mainly rely on:

[0003] Fixed threshold power detection: This method determines the presence of a signal by setting a fixed energy threshold. However, in complex urban electromagnetic environments, this method suffers from a very high false alarm rate due to background noise fluctuations and interference from numerous legitimate wireless devices such as Wi-Fi and Bluetooth.

[0004] Supervised learning based on modulation feature comparison: This method utilizes predefined modulation features such as OOK, FHSS, OFDM, or labeled UAV RF fingerprint databases to train deep learning models such as CNN / ResNet for binary classification or multi-model identification. Although it has high accuracy for known models, it relies heavily on prior knowledge and cannot identify "unknown UAVs" using new protocols, new modulation methods, or encrypted signals, resulting in significant blind spots.

[0005] Static template comparison: requires manual setting of spectrum templates or rule bases, updates are slow, and it cannot cope with agile frequency hopping signals.

[0006] Therefore, the above methods have significant drawbacks: First, they heavily rely on labeled training sets and lack the ability to detect drone signals with unknown protocols or unregistered frequency bands, making them unable to cope with the ever-evolving drone threat and exhibiting weak generalization ability. Second, the fixed threshold method is susceptible to non-stationary noise and background clutter, leading to a significant increase in false negative and false positive rates in low signal-to-noise ratio or high clutter environments, resulting in poor environmental robustness. Third, existing methods mostly employ simple energy peak scanning to locate signals, making it difficult to generate continuous and stable time-frequency trajectories for frequency-hopping signals or signals with instantaneous fading, thus failing to provide reliable targets for subsequent precise jamming. Summary of the Invention

[0007] To address at least one of the aforementioned technical problems, this invention provides a method for detecting unmanned aerial vehicle (UAV) spectrum anomalies and generating trajectories based on unsupervised learning.

[0008] In a first aspect, the present invention provides a method for unsupervised learning-based detection of unmanned aerial vehicle (UAV) spectrum anomalies and trajectory generation, the method comprising:

[0009] Broadband radio frequency signals are continuously acquired using software-defined radio (SDR) to obtain an IQ data stream; a short-time Fourier transform is then performed on the IQ data stream to obtain a time-spectrum diagram.

[0010] The time-frequency spectrum is robustly normalized to obtain a normalized spectrum after suppressing slow drift; constant false alarm rate (CFAR) detection is performed on the normalized spectrum to generate a primary anomaly candidate binary mask.

[0011] Identify spectral sub-blocks corresponding to the primary anomaly candidate binary mask from the standardized spectral map, input the spectral sub-blocks into a pre-trained unsupervised deep learning model to obtain an anomaly scoring map; fuse the anomaly scoring map with the primary anomaly candidate binary mask to generate an anomaly heatmap;

[0012] Density clustering is performed on the abnormal points of the abnormal heatmap in the time-frequency dimension, and the clusters are correlated and filtered across time windows to output the time-frequency trajectory of the UAV signal.

[0013] Preferably, the spectral sub-blocks are input into a pre-trained unsupervised deep learning model to obtain an anomaly scoring map, including:

[0014] The spectral sub-blocks are input into the Variational Autoencoder Generative Adversarial Network (VAE-GAN) model;

[0015] The encoder of the VAE-GAN model maps the spectral sub-blocks to the latent space, and then the decoder reconstructs them, calculating the reconstruction error and Kullback-Leibler divergence.

[0016] Anomaly scores are generated by combining the reconstruction error with the Kullback-Leibler divergence calculation results.

[0017] Preferably, the constant false alarm rate (CFAR) detection includes taking the sorted threshold on a reference window without a target using the order statistic CFAR, and outputting a primary anomaly candidate binary mask.

[0018] The anomaly scoring map is fused with the primary anomaly candidate binary mask to generate an anomaly heatmap. This includes fusing the anomaly scoring map with the primary anomaly candidate binary mask by weighted fusion or conditional random field (CRF) fusion.

[0019] Preferably, the abnormal points in the abnormal heatmap are density-clustered along the time-frequency dimension, and the clusters are correlated and filtered across time windows, including:

[0020] The DBSCAN algorithm is used to divide the outliers into several clusters, resulting in clusters.

[0021] Based on the clusters, Kalman filtering or particle filtering algorithms are used to interpolate and smooth the interrupted trajectories, outputting center frequency, instantaneous bandwidth and activity information with time continuity.

[0022] The interruption trajectory is generated due to signal frequency hopping or instantaneous fading.

[0023] Preferably, after the time-frequency trajectory of the output UAV signal, the method further includes:

[0024] The orientation angle or optical acoustic features of the target within the same time window are obtained by using a DoA array or sensor, and the spatial consistency probability is calculated with the time-frequency trajectory to verify the reliability of the trajectory.

[0025] A feature vector is constructed for each verified trajectory. This feature vector is then input into a pre-trained classification model to obtain a threat score, and the trajectories are sorted according to the magnitude of the threat scores.

[0026] The feature vector includes at least two of the following: duration, active duty cycle, frequency offset stability, bandwidth stability, frequency hopping rhythm rarity, similarity to historical events, DoA consistency, distance to the guard band, and fingerprint residual of neighboring devices.

[0027] The classification model is a lightweight multilayer perceptron (MLP) or logistic regression model, and the threat score threshold is adaptively set using extreme value theory (EVT).

[0028] Preferably, performing a short-time Fourier transform (STFT) on the IQ data stream includes:

[0029] The IQ data stream is divided into several IQ data blocks, and windowing and short-time Fourier transform (STFT) are performed on each IQ data block on a field-programmable gate array (FPGA).

[0030] Robust normalization of the time spectrum includes standardizing the time spectrum using a robust normalization method based on the median m and the absolute median difference MAD.

[0031] Preferably, the VAE-GAN model is trained unsupervised using environmental background and legitimate communication signal samples, and the training objective function is the sum of reconstruction loss, Kullback-Leibler divergence, and adversarial loss.

[0032] Secondly, the present invention also provides a UAV spectrum anomaly detection and trajectory generation system based on unsupervised learning, the system comprising:

[0033] The data acquisition unit is used to continuously acquire broadband radio frequency signals through software-defined radio (SDR) to obtain an IQ data stream; and to perform a short-time Fourier transform on the IQ data stream to obtain a time-spectrum diagram.

[0034] The data processing unit is used to perform robust normalization processing on the time-frequency spectrum to obtain a normalized spectrum after suppressing slow drift; and to perform constant false alarm rate (CFAR) detection on the normalized spectrum to generate a primary anomaly candidate binary mask.

[0035] An anomaly scoring unit is used to identify spectral sub-blocks corresponding to the primary anomaly candidate binary mask from the standardized spectral map, input the spectral sub-blocks into a pre-trained unsupervised deep learning model to obtain an anomaly scoring map, and fuse the anomaly scoring map with the primary anomaly candidate binary mask to generate an anomaly heatmap.

[0036] The trajectory output unit is used to perform density clustering on the abnormal points of the abnormal heatmap in the time-frequency dimension, perform trajectory association and filtering on the clusters across time windows, and output the time-frequency trajectory of the UAV signal.

[0037] Thirdly, the present invention also provides an electronic device including a processor and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein when the processor executes the computer instructions, the electronic device performs the method as described in the first aspect above and any possible implementation thereof.

[0038] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor of an electronic device, cause the processor to perform a method as described in the first aspect above and any possible implementation thereof.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] 1) By employing software-defined radio (SDR) to continuously acquire broadband radio frequency signals and obtain IQ data streams, the limitations of traditional hardware acquisition in terms of bandwidth and flexibility are overcome. This enables comprehensive capture of UAV radio frequency signals in complex electromagnetic environments, avoiding signal omissions due to limited acquisition range. Subsequently, short-time Fourier transform is used to convert the IQ data stream into a time-frequency spectrum, clearly presenting the signal's distribution characteristics in the time-frequency dimension, providing an intuitive and analyzable data format for subsequent anomaly identification. To address the potential slow-varying drift problem in the time-frequency spectrum, robust normalization processing is introduced. This eliminates spectrum baseline drift caused by hardware noise, environmental electromagnetic interference, and other factors during signal acquisition, ensuring that the standardized spectrum more accurately reflects the time-frequency characteristics of the UAV signal. This reduces the impact of interference factors on subsequent anomaly detection results and improves the reliability and accuracy of data preprocessing.

[0041] 2) By performing constant false alarm rate (CFAR) detection on a standardized spectrum to generate a primary anomaly candidate binary mask, it is possible to quickly identify suspected areas where UAV signals may exist based on statistical characteristics. While ensuring a certain level of detection coverage, it efficiently eliminates a large number of obviously anomaly-free spectrum areas, reducing the amount of data required for subsequent deep analysis and improving the overall efficiency of anomaly detection. Simultaneously, it provides a clear analytical object for accurate anomaly identification. The spectrum sub-blocks corresponding to the primary anomaly candidates are input into a pre-trained unsupervised deep learning model. This model does not rely on a large number of labeled UAV signal samples and can accurately distinguish UAV signals from other interference signals such as civilian communication signals and environmental noise by autonomously learning the potential features of the spectrum data. It effectively identifies weak UAV signals or misjudged interference areas that may be missed by CFAR detection, significantly improving the accuracy of anomaly identification and reducing the risk of false positives and false negatives due to the complexity and variability of signals. The anomaly score map output by the unsupervised model is fused with the binary mask of the primary anomaly candidate obtained by CFAR detection to generate an anomaly heatmap. This combines the advantages of the two detection methods, retaining the efficient screening capability of CFAR detection and incorporating the accurate recognition characteristics of the deep learning model. The complementary fusion reduces the limitations of a single detection method, making the judgment of anomaly regions more comprehensive and accurate, and clearly presenting the anomaly distribution of UAV signals in the time and frequency dimensions, providing a reliable basis for subsequent trajectory generation.

[0042] 3) Density clustering of anomaly points in the anomaly heatmap along the time-frequency dimension enables automatic differentiation of anomaly clusters corresponding to different UAV signals based on the degree of clustering. This avoids misclassifying multiple UAV signals as a single signal or splitting intermittent signals from the same UAV into multiple signals, accurately extracting the core anomaly region of each UAV signal, providing clear clustering units for trajectory association, and reducing trajectory fragmentation caused by signal discreteness. By performing trajectory association and filtering across time windows on the clusters, the time-frequency characteristics of clusters within different time windows can be utilized to effectively connect signal clusters of the same UAV within continuous time periods, forming a complete time-frequency trajectory. Simultaneously, filtering can eliminate anomaly clusters caused by instantaneous interference and detection errors, eliminating noise points in the trajectory, making the output UAV signal time-frequency trajectory smoother and more continuous, accurately reflecting the dynamic changes of the UAV in the time-frequency dimension, achieving effective dynamic tracking of UAV signals, and providing accurate trajectory data support for subsequent UAV positioning, behavior analysis, and other applications.

[0043] 4) Employing unsupervised deep learning models eliminates the need for extensive manpower and resources to label UAV signal samples. It can quickly adapt to the signal characteristics of different types of UAVs and signal variations under different electromagnetic environments, solving the problem of poor adaptability of traditional supervised learning models in situations with scarce samples and changing scenarios. This enhances the technology's application capabilities in complex real-world scenarios. From signal acquisition, preprocessing, anomaly detection, and trajectory generation, a complete technology chain is formed. Each stage is technologically mature and tightly integrated. The adopted technologies, such as SDR acquisition, short-time Fourier transform, CFAR detection, unsupervised learning, and density clustering, all have feasible hardware and software support, eliminating the need for specialized and expensive equipment. This facilitates engineering deployment and practical application, making it widely applicable in scenarios requiring UAV control, such as airports, military restricted areas, and large event venues, providing strong technical support for UAV spectrum monitoring and countermeasures.

[0044] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the background art, the accompanying drawings used in the embodiments of the present invention or the background art will be described below.

[0046] Figure 1 A flowchart illustrating a method for detecting unsupervised learning-based UAV spectrum anomalies and generating trajectories, provided in an embodiment of the present invention.

[0047] Figure 2 This is a schematic diagram of a UAV spectrum anomaly detection and trajectory generation system based on unsupervised learning, provided in an embodiment of the present invention. Detailed Implementation

[0048] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0050] Please see Figure 1 , Figure 1 This is a flowchart illustrating a method for detecting unsupervised learning-based UAV spectrum anomalies and generating trajectories, provided in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0051] S10. Continuously acquire broadband radio frequency signals using software-defined radio (SDR) to obtain IQ data streams; perform short-time Fourier transform on the IQ data streams to obtain time-frequency spectrograms;

[0052] Select software-defined radio (SDR) devices that support broadband signal reception, such as USRP or RTL-SDR. Set the SDR's receiving bandwidth according to the UAV's commonly used radio frequency bands, such as the 2.4GHz or 5.8GHz industrial-grade bands, or specific communication bands. Typically, set it to 10MHz-100MHz to ensure coverage of the UAV signal, center frequency, and sampling rate within the target frequency band. The sampling rate must satisfy the Nyquist sampling theorem and is generally no less than twice the highest frequency of the signal. For example, if the highest frequency of the target band is 5.8GHz, the sampling rate should be set to 12GSps or higher.

[0053] Using the signal acquisition software that comes with the SDR, such as GNU Radio or MATLAB's SDR support tools, the SDR device can be controlled to continuously receive broadband radio frequency signals from the surrounding environment. The received analog radio frequency signals are then processed by down-conversion and analog-to-digital converter (ADC) to output an IQ data stream containing signal amplitude and phase information. This data stream is stored in binary file or real-time stream format. The acquisition duration can be set according to monitoring requirements, such as 10-30 seconds for a single acquisition, and continuous loop acquisition is supported.

[0054] The acquired IQ data stream is processed by framing: continuous IQ data is divided into multiple data frames of fixed frame lengths, such as 256 points or 512 points. The frame length needs to balance time and frequency resolution; the longer the frame length, the higher the frequency resolution and the lower the time resolution. The data is adjusted according to the characteristics of the UAV signal, and an overlap rate of 50%-75% is set between adjacent frames to avoid information loss between frames. A window function, such as a Hanning window or a Hamming window, is applied to each data frame to suppress spectral leakage. Then, a Fast Fourier Transform is performed on each windowed frame to convert the time-domain data into frequency-domain data, obtaining the frequency components and corresponding amplitude values ​​of each frame. The time is plotted on the horizontal axis, with the frame number corresponding to the time. For example, with a frame length of 512 points and a sampling rate of 12 GSps, the time of a single frame is approximately 42.67 ns. The horizontal axis time interval is calculated based on the actual time step after frame overlap, the vertical axis is the frequency, the vertical axis range is the frequency interval corresponding to the SDR receiving bandwidth, and the amplitude value is the color depth. The larger the amplitude, the darker the color. A two-dimensional time spectrum is constructed to intuitively present the distribution and changes of the signal in the time-frequency dimension.

[0055] SDR's broadband acquisition capability breaks through the limitations of traditional fixed frequency bands in hardware, covering commonly used communication and navigation frequency bands for UAVs. This avoids signal loss due to frequency band omissions, making it particularly suitable for the synchronous acquisition of signals from multiple types of UAVs in complex electromagnetic environments. STFT converts the one-dimensional IQ data stream into a two-dimensional time-spectrum graph, clearly presenting key characteristics of UAV signals such as frequency jumps, durations, and amplitude changes. For example, it reveals the continuous frequency occupancy characteristics of UAV image transmission signals and the pulse-like frequency characteristics of remote control signals, providing an intuitive "signal profile" for subsequent anomaly identification and solving the problem of time-domain data being difficult to distinguish signal types.

[0056] S20. The time-frequency spectrum is robustly normalized to obtain a normalized spectrum after suppressing slow drift; constant false alarm rate (CFAR) detection is performed on the normalized spectrum to generate a primary anomaly candidate binary mask.

[0057] Preferably, performing a short-time Fourier transform (STFT) on the IQ data stream includes:

[0058] The IQ data stream is divided into several IQ data blocks, and windowing and short-time Fourier transform (STFT) are performed on each IQ data block on a field-programmable gate array (FPGA).

[0059] Robust normalization of the time spectrum includes standardizing the time spectrum using a robust normalization method based on the median m and the absolute median difference MAD.

[0060] An FPGA is essentially a programmable hardware circuit containing a large number of logic cells and configurable routing resources. Designed using a hardware description language, independent hardware resources can be allocated to the processing of each IQ data block, enabling true parallel processing. Multiple data blocks can be processed simultaneously in different areas of the FPGA, resulting in throughput far exceeding that of a sequentially executing general-purpose processor.

[0061] The statistical characteristics of all pixels in the time-frequency spectrum are calculated, with each pixel corresponding to the amplitude value at a specific time-frequency point. The median absolute deviation (MAD) is used instead of the traditional mean-standard deviation statistics to avoid interference from extreme outliers. First, the median M of all pixel amplitude values ​​is calculated, then the absolute deviation of each pixel from the median is calculated, and the median MAD of this absolute deviation is taken. The MAD is multiplied by 1.4826 so that the MAD is equivalent to the standard deviation under a normal distribution to obtain a robust standard deviation σ. Then, the amplitude value of each pixel in the time-frequency spectrum is normalized according to the formula (X - M) / σ, where X is the original pixel amplitude value. The normalized pixel value is mapped to the interval [-3, 3]. Extreme values ​​outside the interval are truncated according to the boundary value to obtain a normalized spectrum. This eliminates the slow drift of the overall spectrum baseline caused by SDR hardware gain fluctuations and changes in environmental noise intensity, such as the slow increase or decrease of the overall spectrum amplitude during long-term acquisition.

[0062] The algorithm selected is either Cell-Averaged CFAR (CA-CFAR) or Ordered Statistical CFAR (OS-CFAR). OS-CFAR has stronger resistance to multi-target interference and is suitable for multi-UAV scenarios. Specifically, the detection window structure is set: with the pixel to be detected as the center, protective cells (to prevent target signal energy from spreading to reference cells, usually set to 8-16 pixels) and reference cells (used to estimate background noise levels, usually set to 32-64 pixels) are set around it. CFAR detection is performed pixel by pixel on the normalized spectrum. The statistical threshold of background noise is calculated using the amplitude values ​​of the reference cells. For example, in CA-CFAR, the mean of the amplitude values ​​of the reference cells is multiplied by the detection threshold coefficient. , Based on the preset false alarm rate setting, such as a false alarm rate of 10⁻ 6 hour Set the threshold to 8-10; if the amplitude value of the detection unit is greater than this threshold, it is judged as a suspected anomaly and marked as 1; otherwise, it is judged as background noise and marked as 0. After traversing all pixels, a binary mask (a two-dimensional matrix containing only 0 and 1) with the same size as the normalized spectrum is generated, which is the primary anomaly candidate binary mask. The position corresponding to "1" in the mask is the suspected UAV signal area.

[0063] In this step, robust normalization effectively suppresses slow spectral drift through statistical methods that resist outliers. This ensures a unified baseline for spectral amplitude across different times and noise environments, preventing baseline drift from causing subsequent detections to misclassify normal noise as abnormal or miss weak anomalous signals, thus improving the consistency and reliability of spectral data. CFAR detection uses statistical thresholds to automate primary screening, controlling the false alarm rate and quickly eliminating non-abnormal background areas. This focuses the analysis on a small number of suspected anomalous regions, significantly reducing the amount of input data for subsequent deep learning models, improving overall detection efficiency, and simultaneously identifying clear targets for deep analysis.

[0064] S30. Identify the spectral sub-blocks corresponding to the primary anomaly candidate binary mask from the standardized spectral map, input the spectral sub-blocks into a pre-trained unsupervised deep learning model to obtain an anomaly scoring map; fuse the anomaly scoring map with the primary anomaly candidate binary mask to generate an anomaly heatmap;

[0065] Based on the position of "1" in the primary anomaly candidate binary mask, a connected component analysis algorithm, such as 8-neighborhood connected component analysis, is used to identify continuous suspected anomaly regions (connected components) in the mask. Each connected component corresponds to a suspected anomaly region in the time-frequency spectrogram. The boundaries of each connected component are expanded. Considering the possibility of weak edge energy in UAV signals, the boundaries of the connected components are expanded outward by 2-4 pixels. The expansion range is set according to the signal bandwidth, forming a rectangular region containing the complete suspected signal. Sub-images (spectral sub-blocks) are extracted from the standardized spectrogram according to the coordinates of this rectangular region. The size of each sub-block is dynamically adjusted according to the size of the anomaly region, for example, 32×32 pixels or 64×64 pixels, and uniformly scaled to the model input size, such as 64×64 pixels.

[0066] An unsupervised model based on an autoencoder (AE) or variational autoencoder (VAE) is selected. During the pre-training stage, a large number of normal time-spectrum sub-blocks without drone signals are used as training data. The model input is the spectrum sub-block. The encoder compresses the sub-block into a low-dimensional feature vector, and the decoder reconstructs the feature vector into a reconstructed sub-block with the same size as the input. The pixel-level reconstruction error between each input spectrum sub-block and the corresponding reconstructed sub-block is calculated, such as the mean squared error (MSE) and the inverse value of the structural similarity index (SSIM). The reconstruction error is used as an anomaly score. The larger the error, the higher the anomaly probability. Each pixel of each sub-block is assigned a corresponding score value according to the anomaly score. The score results of all sub-blocks are concatenated (the score of non-sub-block regions is set to 0) to obtain an anomaly score map with the same size as the original time-spectrum.

[0067] Furthermore, a weighted fusion strategy is adopted, and the primary anomaly candidate binary mask is set as follows: ( This indicates a suspected abnormality. (Indicates normal), abnormal scoring chart is as follows ,in The larger the value, the more significant the anomaly. The fusion formula is: , The weighting coefficient is usually set as follows: This ensures that, based on the suspected areas identified by the mask screening, areas with high anomaly scores are highlighted. The fused area... The values ​​are mapped to the grayscale range of [0, 255] to generate an anomaly heatmap. The higher the grayscale value, the higher the credibility of the anomaly. The high grayscale area in the heatmap is the high credibility drone signal area that has been double-verified.

[0068] Unsupervised models identify differences between normal and abnormal signals through reconstruction errors, effectively distinguishing drone signals from interference signals such as civilian communications and environmental noise. They are particularly effective at capturing weak drone signals that are easily missed by CFAR detection, such as low-power signals from long-range drones, through significant reconstruction errors, thus improving anomaly identification accuracy. The fusion process combines the efficient filtering of CFAR with the precise judgment of deep learning, avoiding false detections caused by noise fluctuations in single CFAR detection and false detections caused by variations in normal signals in a single deep learning model. This makes high-confidence areas in the anomaly heatmap more closely resemble real drone signals, providing accurate signal location information for subsequent trajectory generation.

[0069] S40. Density clustering is performed on the abnormal points of the abnormal heatmap in the time-frequency dimension, and the clusters are correlated and filtered across time windows to output the time-frequency trajectory of the UAV signal.

[0070] First, threshold segmentation is performed on the abnormal heatmap. A grayscale threshold is set, and pixels in the heatmap with grayscale values ​​exceeding the threshold are extracted as abnormal points. The coordinates of each abnormal point are represented as follows: , Using time as the coordinate, Using frequency coordinates, a set of outliers is formed. .

[0071] The density-based spatial clustering algorithm DBSCAN is used to cluster set P, with two key parameters set: neighborhood radius. Based on the temporal continuity and frequency bandwidth of the drone signal, such as the time dimension... Frame, frequency dimension ,comprehensive The threshold and minimum number of points (MinPts) corresponding to the Euclidean distance are used to ensure that the clusters contain a sufficient number of outliers, avoiding the formation of false clusters from isolated noise points. The outlier set is traversed, and outliers with a density reachable from the core point (points in their neighborhood containing ≥MinPts outliers) are grouped into the same cluster, ultimately resulting in multiple clusters. Each cluster corresponds to a potential drone signal.

[0072] Set time window division rules: Divide the entire data collection period into multiple consecutive time windows of fixed duration. Each time window contains clusters corresponding to that time period. To calculate the feature similarity of clusters within adjacent time windows, the center coordinates of each cluster are first extracted. , , This represents the time mean and frequency mean of all points within the cluster. It also determines the cluster size. The standard deviation of time and frequency within a cluster is used to calculate similarity using weighted Euclidean distance. If the calculated result exceeds the set similarity threshold, it is determined to be a continuous cluster of signals from the same UAV, and an association relationship is established.

[0073] Finally, Kalman filtering or moving average filtering is applied to the preliminary trajectory formed by association (a broken line composed of the coordinates of the cluster centers of multiple time windows): the time and frequency coordinates of the trajectory are used as state variables, and a state transition model is established based on the motion characteristics of UAV signals, such as smooth frequency changes and continuous time. By filtering, the jump points in the trajectory, such as the cluster center offset caused by instantaneous interference, are eliminated, and a smooth continuous trajectory is obtained. Finally, the time-frequency trajectory of each UAV signal is output, which is a continuous curve with time as the horizontal axis and frequency as the vertical axis.

[0074] DBSCAN clustering can automatically identify anomalous point clusters of different densities, effectively distinguishing signals from multiple drones within the same time period. This avoids misclassifying multiple target signals as a single signal, resolving signal confusion issues in scenarios where multiple drones coexist. Cross-time window association ensures effective connection of signal clusters from the same drone within a continuous time period, preventing trajectory breaks caused by signal loss due to brief signal interruptions, such as drone obstruction. Filtering eliminates trajectory noise, ensuring that the output time-frequency trajectory accurately reflects the dynamic changes of the drone, such as frequency fine-tuning of remote control signals and continuous frequency occupancy of image transmission signals during drone flight. This provides precise trajectory data support for drone positioning, flight path analysis, and abnormal behavior warnings.

[0075] In one embodiment, the spectral sub-blocks are input into a pre-trained unsupervised deep learning model to obtain an anomaly scoring map, including:

[0076] The spectral sub-blocks are input into the Variational Autoencoder Generative Adversarial Network (VAE-GAN) model;

[0077] The encoder of the VAE-GAN model maps the spectral sub-blocks to the latent space, and then the decoder reconstructs them, calculating the reconstruction error and Kullback-Leibler divergence.

[0078] Anomaly scores are generated by combining the reconstruction error with the Kullback-Leibler divergence calculation results.

[0079] Before system deployment, the VAE-GAN model needs to be unsupervised pre-trained on a large dataset containing only "normal" background spectrum sub-blocks. The model structure is as follows:

[0080] The encoder is a neural network that takes a spectral sub-block image x and outputs two vectors in the latent space: the mean and the variance. This represents a Gaussian distribution.

[0081] Decoder: A neural network that samples a point z from the latent space and reconstructs it into a spectral map x' of the same size as the original input x.

[0082] Discriminator: A neural network that receives a spectral sub-block (which can be a real normal sample x or a reconstructed sample x' generated by the decoder) and determines whether it is "real" or "generated".

[0083] The training process is an adversarial game between the encoder / decoder (generator) and the discriminator, while the VAE part aims to regularize the latent distribution. The VAE loss encourages the reconstructed sample x' to be similar to the input x (reconstruction error) and encourages the distribution of the latent variables to approximate a standard normal distribution. The GAN loss encourages the discriminator to better distinguish between real and generated data, while also encouraging the decoder to generate data sufficient to "fool" the discriminator. The final total loss function is a weighted sum of the VAE loss and the GAN loss.

[0084] In the online detection phase, spectral sub-blocks are input into the Variational Autoencoder Generative Adversarial Network (VAE-GAN) model. First, the spectral sub-block is input to the encoder to obtain the mean and variance of its latent distribution parameters. The decoder then reconstructs the sub-block to obtain the output. The reconstruction error measures the model's ability to "copy" or "remember" the current input sub-block. For normal samples within the training domain, the model reconstructs them well; however, for unknown anomalous samples such as drone signals, the reconstruction error can be large. The mean squared error (MSE) or mean absolute error is typically used to calculate the reconstruction error.

[0085] The Kullback-Leibler divergence measures the deviation of the latent distribution of test samples from the standard normal prior distribution. During training, the model is constrained to map all normal samples into the latent space. An anomalous sample may be mapped to an uncommon region in the latent space, off-center from the origin. The "exotic" characteristics of anomalous samples cause their latent encoding to differ from the concentrated distribution of normal samples, resulting in a larger Kullback-Leibler divergence value. The final anomalous score is a weighted joint of the two losses mentioned above. Hyperparameters, or weights, are used to balance the importance of the two metrics. Optimal weights can usually be determined by adjusting them on the validation set, or simply by normalizing and adding them together. For a spectral sub-block to be considered normal, it must simultaneously meet two conditions: it must be well reconstructed by the model, and its latent encoding must conform to the distribution of normal data. Any violation of either condition will lead to a high anomalous score. Drone signals, as a significant anomalous signal, typically cause both increased reconstruction error and increased KL divergence.

[0086] For each candidate spectrogram sub-block provided by CFAR, the above steps are repeated to calculate a corresponding joint anomaly score. All scores are then mapped back to their corresponding positions in the original temporal spectrogram, thus forming an anomaly score map. The value of each pixel in the map represents the probability of an anomalous signal existing at that location; the higher the score, the greater the probability.

[0087] Compared to autoencoders that rely solely on reconstruction error, VAE-GAN adds a distributional constraint through KL divergence. Even if an anomalous signal is reconstructed relatively well by chance, as long as its features are fundamentally different from normal data, its latent coding distribution will be anomalous, thus being detected and avoiding false negatives. The introduction of GANs typically helps decoders generate clearer and more realistic reconstructed images, which contributes to more accurate and stable calculation of reconstruction error. The latent space of VAEs is continuous and structured (following a standard normal distribution), giving "normal" and "abnormal" more explicit geometric meanings in the latent space (such as distance from the origin), which is beneficial for improving the interpretability and discriminative power of anomaly scoring.

[0088] Preferably, the constant false alarm rate (CFAR) detection includes taking the sorted threshold on a reference window without a target using the order statistic CFAR, and outputting a primary anomaly candidate binary mask.

[0089] The anomaly scoring map is fused with the primary anomaly candidate binary mask to generate an anomaly heatmap. This includes fusing the anomaly scoring map with the primary anomaly candidate binary mask by weighted fusion or conditional random field (CRF) fusion.

[0090] On the standardized spectrum, a detection threshold is adaptively set so that the probability of false alarms remains constant regardless of changes in background noise. OS-CFAR is particularly good at maintaining stable performance in scenarios with multiple targets or clutter edges.

[0091] Sliding window processing involves a detection window sliding pixel-by-pixel or cell-by-cell across a normalized spectral map. For the current cell to be detected at the center of the window, this includes:

[0092] Define a reference window: Define a reference sliding window around the current cell that does not contain any guard cells. This reference window includes a group of reference cells near the current cell, but not so much that strong signals would leak in and affect the background estimation.

[0093] Sorting and Selection: Sort the signal strength values ​​of all reference cells from smallest to largest to form an ordered list.

[0094] Threshold calculation: Instead of taking the average of all reference cells (such as CA-CFAR), the Kth value in this ordered list is selected as the estimate of the current background noise level. This K value is a key parameter that determines the characteristics of the algorithm. For example, choosing the median is very insensitive to outliers.

[0095] Threshold setting and judgment: The selected Kth value is multiplied by a scaling factor, which is calculated offline based on the expected false alarm rate, to obtain the final detection threshold. If the intensity value of the central unit to be detected exceeds this threshold, it is determined as a primary anomaly candidate point and marked as 1 (white) at the corresponding position in the binary mask; otherwise, it is marked as 0 (black).

[0096] Because sorted values ​​are selected, even if a few other strong interference signals, such as another weak UAV signal or sudden noise, are mixed into the reference window, these "outliers" will not significantly increase the threshold estimate, ensuring that strong targets mask weak targets. This greatly improves the ability to detect multiple weak targets in complex electromagnetic environments.

[0097] The fusion of the anomaly score map and the binary mask to generate an anomaly heatmap mainly combines the previous CFAR detection results (either 0 or 1) with the anomaly score (a continuous probability value) given by the deep learning model. This generates a refined heatmap that includes both spatial location information and anomaly confidence information, providing optimal input for the final trajectory generation. Weighted fusion or Conditional Random Field (CRF) fusion can be used. Weighted fusion essentially amplifies and highlights the anomaly score from deep learning within the region of interest defined by CFAR, resulting in higher scores for genuine signal points and relatively lower scores for accidental noise points. CRF fusion uses both the initial anomaly candidate binary mask and the anomaly score map as input observation data to the CRF model.

[0098] Furthermore, the energy function is established: The CRF model defines an energy function, which consists of two parts:

[0099] Unary energy: Based on the observation data of each pixel itself. For example, if a point has a high anomaly score in a deep learning model, then its unary energy for being labeled as "anomaly" is low (low cost). CFAR masks can also serve as strong priors for unary energy; if a point is not detected by CFAR, then its energy for being labeled as "background" is very low.

[0100] Binary energy: This is the core of CRF. It considers the relationship between adjacent pixels. A common principle is that adjacent pixels with similar spectral features (such as intensity) are more likely to belong to the same category (both anomalies or both background). Assigning them different labels results in a high penalty energy.

[0101] Inference and Optimization: The goal of the CRF model is to find a final label map (i.e., a heatmap, but obtained through optimization) that minimizes the total energy function described above. This optimization process forces the result to become spatially smooth: isolated low-intensity points surrounded by high-intensity outliers may be "lifted" (considered as part of the signal); conversely, scattered high-intensity points surrounded by low-intensity regions may be "suppressed" (considered as noise).

[0102] Thus, the above embodiments can filter out isolated noise points and fill in gaps and voids in the actual signal trajectory that may occur due to occasional detection failures, forming a more continuous and complete signal region. It not only relies on information from individual points but also utilizes information from surrounding points for comprehensive decision-making, resulting in more robust and accurate results.

[0103] Preferably, the abnormal points in the abnormal heatmap are density-clustered along the time-frequency dimension, and the clusters are correlated and filtered across time windows, including:

[0104] The DBSCAN algorithm is used to divide the outliers into several clusters, resulting in clusters.

[0105] Based on the clusters, Kalman filtering or particle filtering algorithms are used to interpolate and smooth the interrupted trajectories, outputting center frequency, instantaneous bandwidth and activity information with time continuity.

[0106] The interruption trajectory is generated due to signal frequency hopping or instantaneous fading.

[0107] The DBSCAN algorithm is used to process points in the anomaly heatmap. The algorithm automatically identifies high-density, clustered regions on a two-dimensional plane defined by time and frequency, recognizing each such region as an independent signal cluster. This effectively groups multiple discrete points representing the same signal source while filtering out isolated points that may be noise. Kalman filtering and other prediction algorithms are employed to address problems caused by signal frequency hopping or brief disappearances (i.e., "track interruptions"). This algorithm intelligently predicts the next possible location of a signal cluster based on its past positions and trends. The re-emerging signal clusters after the interruption are matched with the predicted locations to determine if they belong to the same track. The track during the interruption is completed, and the entire track is smoothed to eliminate unnecessary jitter. Finally, time-continuous, smoothly changing tracks are generated, and key information such as how the center frequency of each track changes over time, signal bandwidth, and active duration are extracted.

[0108] In one embodiment, following the time-frequency trajectory of the output drone signal, the method further includes:

[0109] The orientation angle or optical acoustic features of the target within the same time window are obtained by using a DoA array or sensor, and the spatial consistency probability is calculated with the time-frequency trajectory to verify the reliability of the trajectory.

[0110] A feature vector is constructed for each verified trajectory. This feature vector is then input into a pre-trained classification model to obtain a threat score, and the trajectories are sorted according to the magnitude of the threat scores.

[0111] The feature vector includes at least two of the following: duration, active duty cycle, frequency offset stability, bandwidth stability, frequency hopping rhythm rarity, similarity to historical events, DoA consistency, distance to the guard band, and fingerprint residual of neighboring devices.

[0112] The classification model is a lightweight multilayer perceptron (MLP) or logistic regression model, and the threat score threshold is adaptively set using extreme value theory (EVT).

[0113] Spatial consistency verification aims to utilize multimodal sensor data to cross-verify the authenticity of the spectral trajectory from a spatial dimension, eliminating interference from spurious signals. Specifically, it includes:

[0114] Multimodal data acquisition: While the core detection process is running, the system simultaneously acquires data through other integrated sensors (such as microphone arrays and optical cameras) or by using the orientation finding (DoA) array composed of the SDR itself.

[0115] Feature extraction: The DoA array processes radio frequency signals within the same time window to estimate the direction of arrival of the signal source and obtain angular information. Optical / acoustic sensors use computer vision or audio processing algorithms to detect whether there is a drone target within the field of view or audible range, and similarly estimate its azimuth angle.

[0116] Spatial consistency calculation: Align the time-frequency trajectory obtained from spectrum analysis with the azimuth trajectory obtained from the sensor on a timeline. Probability calculation: Evaluate the degree of consistency between these two independently generated trajectories in spatial direction. For example, calculate whether the direction of the signal source matches the direction of the drone detected by the optical sensor within the same time period. Finally, generate and output a spatial consistency confidence score. The higher the score, the greater the probability that the spectrum trajectory was generated by a real drone, thus verifying its reliability. Only trajectories that pass verification (confidence exceeding a certain threshold) will proceed to the next stage of threat assessment.

[0117] Threat assessment and ranking aims to intelligently analyze verified drone trajectories, automatically assess their threat levels, prioritize alerts, and optimize the scheduling of response resources. When constructing multi-dimensional feature vectors, the system extracts a set of features that characterize the behavioral patterns of each verified trajectory, building a feature vector. These features include, but are not limited to:

[0118] Duration: How long the signal lasted in total. The longer the duration, the clearer the intention may be.

[0119] Active duty cycle: The proportion of signal active time to total duration. A high duty cycle may indicate that image transmission or continuous control is in progress.

[0120] Frequency offset stability / bandwidth stability: the degree of jitter in the signal center frequency and bandwidth. Poor stability may indicate a low-quality or homemade drone platform.

[0121] Frequency hopping rhythm rarity: Whether the frequency hopping pattern of the signal is rare. Rare patterns may correspond to specific models or non-commercial drones.

[0122] Similarity to historical events: The degree of similarity between the characteristics of the current trajectory and previously recorded malicious drone intrusion events.

[0123] DoA Consistency: Confidence score from the previous spatial validation step.

[0124] Distance from protected frequency bands: Whether the signal frequency is close to protected critical communication frequency bands such as those of airports and government offices.

[0125] Nearby device fingerprint residual: The degree of difference between subtle signal features (RF fingerprint) and known friendly devices in the database, used to identify unknown devices.

[0126] In practice, not all features are needed. Depending on the application scenario and computing resources, at least two or more of the most discriminative features can be selected and combined. For threat scoring, a computationally efficient lightweight multilayer perceptron or logistic regression model is used. These models are pre-trained using historical data (labeled with normal flight, suspicious flight, malicious flight, etc.). The constructed feature vector is input into this classification model. The model outputs a continuous threat score (e.g., a value between 0 and 1), with higher scores indicating greater threats. Traditional fixed thresholds are difficult to adapt to changes in different environments. Therefore, this embodiment uses extreme value theory to dynamically set the alarm threshold. EVT estimates the probability of "extreme outliers" occurring in the current environment by analyzing the "tails" (i.e., extremely high values) of the historical threat score distribution, thereby adaptively determining a reasonable threshold. This allows the system to better cope with baseline changes in threat levels under different scenarios. The system sorts all trajectories from high to low according to their threat scores and selects those exceeding the EVT threshold to generate the final alarm. This allows security personnel to prioritize and handle the most threatening and suspicious targets, greatly improving emergency response efficiency.

[0127] This embodiment provides a complete "detection-verification-evaluation-alarm" closed loop. It improves reliability through multi-sensor fusion, achieves intelligent threat understanding through multi-dimensional feature analysis, and finally realizes efficient and flexible decision output through lightweight models and adaptive thresholds.

[0128] Preferably, the VAE-GAN model is trained unsupervised using environmental background and legitimate communication signal samples, and the training objective function is the sum of reconstruction loss, Kullback-Leibler divergence, and adversarial loss.

[0129] The training data uses clean environmental background noise and legitimate communication signal samples. Crucially, it contains absolutely no drone signals. VAE-GAN is a hybrid model combining the advantages of Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs). The VAE part learns how to compress a normal spectrogram into a compact, structured "code" (latent vector) and then reconstruct the original image from this code as accurately as possible. It is constrained so that the codes of all normal data follow a simple standard normal distribution. The GAN part simultaneously trains a discriminator tasked with distinguishing between "real" normal spectrograms and the reconstructed spectrograms "generated" by the VAE. This forces the VAE's decoder (acting as a generator) to produce increasingly realistic reconstructed images. Joint training enables the model to learn and represent patterns in the "normal" spectrum with extremely high accuracy.

[0130] During training, normal spectrogram sub-blocks are input into the model, and the encoder maps them to a distribution (mean and variance) in the latent space. The decoder samples from this distribution and reconstructs the spectrogram. The discriminator receives both the real and reconstructed spectrograms and attempts to distinguish between them. The model parameters are updated by minimizing a comprehensive objective function, which is a weighted sum of three terms: Reconstruction loss measures the pixel-level difference between the reconstructed and original spectrograms, ensuring the model accurately "remembers" normal patterns. KL divergence measures the difference between the distribution of latent variables and a standard normal distribution, ensuring the latent space is regular and ordered, facilitating sampling and interpolation. Adversarial loss encourages the decoder to generate photorealistic reconstructed images that fool the discriminator. This improves the quality of the reconstructed images, making anomalous signals easier to detect.

[0131] The above embodiments eliminate the need to collect and label difficult-to-obtain signal data from various drone models, solving the pain points of sample scarcity and labeling difficulties, and significantly reducing deployment barriers and costs. Because the model learns from a "normal" benchmark, it can detect novel or even unknown modulation schemes of drone signals, demonstrating excellent generalization capabilities. The introduction of GANs makes the reconstructed normal background signals very clear and realistic. This makes anomalous signals stand out more in comparison and easier to identify. It can not only accurately detect known drone threats but also discover unknown new threats, and it is simple to deploy and inexpensive.

[0132] See Figure 2 The present invention also provides a system for detecting spectral anomalies and generating trajectories of unsupervised drones based on unsupervised learning, the system comprising:

[0133] Data acquisition unit 100 is used to continuously acquire broadband radio frequency signals through software-defined radio (SDR) to obtain IQ data streams; and to perform short-time Fourier transform on the IQ data streams to obtain time-spectrum diagrams.

[0134] The data processing unit 200 is used to perform robust normalization processing on the time-frequency spectrum to obtain a normalized spectrum after suppressing slow drift; and to perform constant false alarm rate (CFAR) detection on the normalized spectrum to generate a primary anomaly candidate binary mask.

[0135] Anomaly scoring unit 300 is used to identify spectral sub-blocks corresponding to the primary anomaly candidate binary mask from the standardized spectral map, input the spectral sub-blocks into a pre-trained unsupervised deep learning model to obtain an anomaly scoring map, and fuse the anomaly scoring map with the primary anomaly candidate binary mask to generate an anomaly heatmap.

[0136] The trajectory output unit 400 is used to perform density clustering on the abnormal points of the abnormal heatmap in the time-frequency dimension, perform trajectory association and filtering on the clusters across time windows, and output the time-frequency trajectory of the UAV signal.

[0137] It is understood that the system provided in this embodiment has functions or includes modules that can be used to execute the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0138] The present invention also provides an electronic device including a processor and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein when the processor executes the computer instructions, the electronic device performs a method as described in any of the above possible implementations.

[0139] The present invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor of an electronic device, cause the processor to perform a method as described in any of the above possible implementations.

[0140] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

Claims

1. A method for detecting spectral anomalies and generating trajectories of unsupervised UAVs based on unsupervised learning, characterized in that, The method includes: Broadband radio frequency signals are continuously acquired using software-defined radio (SDR) to obtain an IQ data stream; a short-time Fourier transform is then performed on the IQ data stream to obtain a time-spectrum diagram. The time-frequency spectrum is robustly normalized to obtain a normalized spectrum after suppressing slow drift; constant false alarm rate (CFAR) detection is performed on the normalized spectrum to generate a primary anomaly candidate binary mask. Identify spectral sub-blocks corresponding to the primary anomaly candidate binary mask from the standardized spectral map, input the spectral sub-blocks into a pre-trained unsupervised deep learning model to obtain an anomaly scoring map; fuse the anomaly scoring map with the primary anomaly candidate binary mask to generate an anomaly heatmap; Density clustering is performed on the abnormal points of the abnormal heatmap in the time-frequency dimension, and the clusters are correlated and filtered across time windows to output the time-frequency trajectory of the UAV signal.

2. The method for UAV spectrum anomaly detection and trajectory generation based on unsupervised learning according to claim 1, characterized in that, The spectral sub-blocks are input into a pre-trained unsupervised deep learning model to obtain an anomaly scoring map, including: The spectral sub-blocks are input into the Variational Autoencoder Generative Adversarial Network (VAE-GAN) model; The encoder of the VAE-GAN model maps the spectral sub-blocks to the latent space, and then the decoder reconstructs them, calculating the reconstruction error and Kullback-Leibler divergence. Anomaly scores are generated by combining the reconstruction error with the Kullback-Leibler divergence calculation results.

3. The method for UAV spectrum anomaly detection and trajectory generation based on unsupervised learning according to claim 1, characterized in that, The execution of constant false alarm rate (CFAR) detection includes using the order statistic CFAR to take the sorted threshold on a reference window without a target, and outputting a primary anomaly candidate binary mask. The anomaly scoring map is fused with the primary anomaly candidate binary mask to generate an anomaly heatmap. This includes fusing the anomaly scoring map with the primary anomaly candidate binary mask by weighted fusion or conditional random field (CRF) fusion.

4. The method for unsupervised learning-based UAV spectrum anomaly detection and trajectory generation according to claim 1, characterized in that, Density clustering is performed on the outliers of the abnormal heatmap along the time-frequency dimension, and trajectory association and filtering are performed on the clusters across time windows, including: The DBSCAN algorithm is used to divide the outliers into several clusters, resulting in clusters. Based on the clusters, Kalman filtering or particle filtering algorithms are used to interpolate and smooth the interrupted trajectories, outputting center frequency, instantaneous bandwidth and activity information with time continuity. The interruption trajectory is generated due to signal frequency hopping or instantaneous fading.

5. The method for unsupervised learning-based UAV spectrum anomaly detection and trajectory generation according to claim 1, characterized in that, Following the time-frequency trajectory of the output drone signal, the following is also included: The orientation angle or optical acoustic features of the target within the same time window are obtained by using a DoA array or sensor, and the spatial consistency probability is calculated with the time-frequency trajectory to verify the reliability of the trajectory. A feature vector is constructed for each verified trajectory. This feature vector is then input into a pre-trained classification model to obtain a threat score, and the trajectories are sorted according to the magnitude of the threat scores. The feature vector includes at least two of the following: duration, active duty cycle, frequency offset stability, bandwidth stability, frequency hopping rhythm rarity, similarity to historical events, DoA consistency, distance to the guard band, and fingerprint residual of neighboring devices. The classification model is a lightweight multilayer perceptron (MLP) or logistic regression model, and the threat score threshold is adaptively set using extreme value theory (EVT).

6. The method for unsupervised learning-based UAV spectrum anomaly detection and trajectory generation according to claim 1, characterized in that, Performing a short-time Fourier transform (STFT) on the IQ data stream includes: The IQ data stream is divided into several IQ data blocks, and windowing and short-time Fourier transform (STFT) are performed on each IQ data block on a field-programmable gate array (FPGA). Robust normalization of the time spectrum includes standardizing the time spectrum using a robust normalization method based on the median m and the absolute median difference MAD.

7. The method for unsupervised learning-based UAV spectrum anomaly detection and trajectory generation according to claim 2, characterized in that, The VAE-GAN model is trained unsupervised using environmental background and legitimate communication signal samples. The training objective function is the sum of reconstruction loss, Kullback-Leibler divergence, and adversarial loss.

8. A system for detecting spectral anomalies and generating trajectories of unsupervised UAVs based on unsupervised learning, characterized in that, The system includes: The data acquisition unit is used to continuously acquire broadband radio frequency signals through software-defined radio (SDR) to obtain IQ data streams; and to perform short-time Fourier transform on the IQ data streams to obtain time-spectrum diagrams. The data processing unit is used to perform robust normalization processing on the time-frequency spectrum to obtain a normalized spectrum after suppressing slow drift; and to perform constant false alarm rate (CFAR) detection on the normalized spectrum to generate a primary anomaly candidate binary mask. An anomaly scoring unit is used to identify spectral sub-blocks corresponding to the primary anomaly candidate binary mask from the standardized spectral map, input the spectral sub-blocks into a pre-trained unsupervised deep learning model to obtain an anomaly scoring map, and fuse the anomaly scoring map with the primary anomaly candidate binary mask to generate an anomaly heatmap. The trajectory output unit is used to perform density clustering on the abnormal points of the abnormal heatmap in the time-frequency dimension, perform trajectory association and filtering on the clusters across time windows, and output the time-frequency trajectory of the UAV signal.

9. An electronic device, characterized in that, include: A processor and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein when the processor executes the computer instructions, the electronic device performs the unsupervised learning-based UAV spectrum anomaly detection and trajectory generation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which includes program instructions that, when executed by a processor of an electronic device, cause the processor to perform the unsupervised learning-based UAV spectrum anomaly detection and trajectory generation method according to any one of claims 1 to 7.

Images