Unmanned aerial vehicle signal detection and identification method based on spectrum feature enhancement

Abstract

The invention discloses an unmanned aerial vehicle signal detection and identification method based on spectrum feature enhancement. The method comprises the following steps: firstly, carrying out time-frequency transformation on a received signal to obtain an original time-frequency graph; performing learning enhancement on the degenerated time-frequency graph by adopting a coding-decoding type deep neural network, and realizing noise suppression and structure recovery by combining pixel reconstruction, structural similarity and a texture perception loss function; and finally, performing target area detection and positioning on the enhanced time-frequency graph, directly outputting a structured result containing a time-frequency range, a category and confidence, and completing conversion from a frequency spectrum to a linkable engineering target. According to the unmanned aerial vehicle signal detection and recognition method based on spectrum feature enhancement, a spectrum feature enhancement mechanism is introduced before traditional spectrum analysis and feature recognition processing, and region-level detection and judgment are executed under the enhanced spectrum constraint condition; reliable discovery, positioning and identification of an unmanned aerial vehicle control link and an image transmission link in a complex electromagnetic environment are realized.

Description

A method for UAV signal detection and recognition based on spectral feature enhancement Technical Field

[0001] This invention belongs to the field of UAV radio detection and early warning technology, and specifically relates to a method for UAV signal detection and identification based on spectrum feature enhancement. Background Technology

[0002] In recent years, with the rapid development of consumer drone technology, small multi-rotor drones have been widely used in personal entertainment, aerial photography, facility inspection, and emergency rescue due to their small size, low cost, ease of operation, and ability to carry high-definition cameras and stable image transmission modules. However, with the rapid increase in the number of drones, their illegal use in sensitive areas such as airports, prisons, power facilities, oil and petrochemical parks, key urban areas, and major event venues has become increasingly prominent. This includes activities such as cross-border aerial photography, privacy theft, delivery of contraband into controlled areas, and long-term reconnaissance and surveillance of key targets, posing new challenges to public safety and the protection of critical infrastructure. Illegal drones are typically small, flexible in deployment, and possess strong communication anonymity, putting significant pressure on traditional security systems.

[0003] To address the problem of unauthorized drone intrusion, counter-UAS (counter-drone) technology has been gradually developed, primarily encompassing radar detection, electro-optical / infrared imaging, acoustic detection, and electromagnetic spectrum-based wireless signal detection. Among these, radar is susceptible to ground clutter and low radar cross-sections in scenarios involving small, low-altitude targets; electro-optical and infrared systems are limited by lighting, weather, and obstruction conditions; and acoustic detection is easily interfered with by background noise in complex environments. In contrast, passive detection methods based on the electromagnetic spectrum, which monitor the wireless communication link between the drone and its remote controller or ground station without actively radiating electromagnetic signals, offer advantages such as good concealment, wide coverage, and all-weather operation, making them an important technological direction for counter-UAS systems.

[0004] However, in practical engineering applications, spectrum detection faces a series of key challenges. Illegal drones often operate in ISM bands such as 2.4GHz and 5.8GHz, sharing spectrum resources with numerous Wi-Fi, Bluetooth, industrial control wireless, and video transmission devices. Their control and transmission links highly overlap in the frequency domain and exhibit sudden and irregular behavior in time. Simultaneously, illegal drones frequently operate at long distances, in non-line-of-sight, or power-constrained environments, where the signal-to-noise ratio of their received signals may drop to -5dB or even lower. Under these conditions, traditional methods based on energy thresholds, spectrum templates, or simple statistical characteristics struggle to reliably highlight drone signals in strong background interference, leading to a significant increase in both false negative and false positive rates.

[0005] To enhance spectrum detection capabilities, deep learning methods have been introduced into the field of UAV signal recognition in recent years. By automatically extracting features from time-frequency maps, they enable the identification of signal types or communication systems. However, existing methods often assume that the input time-frequency map has a high signal-to-noise ratio (SNR) and clear texture structure, typically relying on "clear spectra" obtained under high-quality sampling conditions. In environments with extremely low SNR, multi-source interference, and highly overlapping spectra, problems such as broken spectral lines, blurred textures, and indistinct edges are common in time-frequency maps. This makes it difficult for deep models to extract stable features, and their detection performance exhibits a significant "cliff-like" drop as the SNR decreases. Simply increasing the sample size or adjusting the network structure cannot fundamentally solve the performance bottleneck caused by insufficient separability of the input spectrum.

[0006] Furthermore, most existing studies directly apply deep models to signal detection or classification, with few specifically designed for the "visibility and structural separability of the spectrum itself" at the system level. There is still a lack of a systematic technical solution for structurally enhancing the original time-frequency map under conditions of extremely low signal-to-noise ratio, multi-source interference, and spectral texture degradation, and on this basis, completing signal region localization and category determination.

[0007] In summary, existing UAV spectrum detection technologies generally suffer from the following shortcomings: reliable extraction of spectral structure is difficult under extremely low signal-to-noise ratio conditions; robustness is insufficient in environments with multi-source interference and highly overlapping spectra; and deep learning methods are highly dependent on the quality of the input spectrum and lack pre-enhancement mechanisms for degraded spectra. Therefore, it is necessary to propose a UAV signal detection and recognition technology scheme that focuses on spectral feature enhancement and combines it with signal region detection and decision-making to achieve stable and engineerable UAV monitoring and early warning capabilities in complex electromagnetic environments. Summary of the Invention

[0008] Purpose of the invention: In order to overcome the above shortcomings, the purpose of this invention is to provide a method for UAV signal detection and identification based on spectrum feature enhancement. By introducing a spectrum feature enhancement mechanism before traditional spectrum analysis and feature recognition processing, and performing regional-level detection and decision under enhanced spectrum constraints, reliable detection, location and identification of UAV control links and image transmission links are achieved in complex electromagnetic environments.

[0009] Technical Solution: To achieve the above objectives, this invention provides a method for UAV signal detection and recognition based on spectral feature enhancement, comprising: S1): signal acquisition and time-frequency transformation, specifically including: S101): preprocessing of the RF receiving link and baseband signal to obtain a steady-state baseband signal; S102): frame-segmentation, windowing, and short-time Fourier transform to obtain a complex spectrum representation to characterize the joint distribution characteristics of the signal in the time and frequency dimensions; S103): power spectrum construction and generation of the original time-frequency map; S2): spectral feature enhancement, specifically including: S201): training sample construction, using a high-quality time-frequency map as a priori spectral structure, with the reference time-frequency map denoted as... ; Perform spectral degradation modeling to generate the corresponding degradation time-frequency graph; S202): Implement forward mapping through a spectral enhancement network structure, assuming the input degradation time-frequency graph is Let the spectrum enhancement network be denoted as... Its output enhancement time-frequency diagram is as follows: The spectrum enhancement network adopts an encoder-decoder deep neural network structure. Through the above structural design, the enhancement network can complete the structural reconstruction of the degraded time-frequency map in a single forward inference process, meeting the requirements of real-time or near-real-time applications. (S203): Loss function design and enhancement objective constraint. A combined optimization objective consisting of multiple losses is adopted, specifically: a pixel-level reconstruction loss is introduced to constrain the consistency of the enhanced time-frequency map and the reference time-frequency map in the overall energy distribution, as shown in the following formula: A structural similarity loss is introduced to constrain spectral edges, spectral line continuity, and local structural morphology, as shown in the following formula: A texture-aware loss is introduced to measure the difference between the enhanced result and the reference result in a high-dimensional feature space through a feature extraction network. This is used to emphasize the texture restoration effect related to signal recognition, as shown in the following formula: The comprehensive loss function is defined as: in The weighting coefficients are used to balance noise suppression strength, structural fidelity, and texture enhancement effect. The spectrum feature enhancement model obtained through the above training makes the enhanced time-frequency map more stable in terms of statistical distribution and structural characteristics, providing reliable input for the subsequent broadband multi-region detection module to perform unified threshold determination and region localization. S3): Signal region detection and target determination based on enhanced spectrum, specifically including: S301): Target detection input and modeling based on enhanced spectrum constraints; S302): Target region localization and generation of an initial candidate signal set; S303): Candidate result confirmation and structured output. Through the above target detection and confirmation mechanism, together with spectrum feature enhancement, a complete processing link of "spectrum enhancement - target discovery - result confirmation - system linkage" is formed, enabling the overall solution to have stable and engineerable UAV signal detection capabilities in complex electromagnetic environments.

[0010] Further, step S101 specifically involves: the received radio frequency signal being sequentially subjected to a pre-selected bandpass filter to suppress strong out-of-band interference, low-noise amplification, and frequency conversion down-conversion processing to obtain a complex baseband analog signal covering the target frequency band; the baseband signal is then processed by an analog-to-digital converter at a sampling rate... Quantization bit width Sampling and quantization are performed to obtain a discrete-time complex sequence: in The quantization operator is used. Considering the common non-ideal factors in actual reception, such as DC bias, automatic gain fluctuations, and I / Q imbalance, baseband preprocessing is performed on the discrete signal before time-frequency analysis to obtain a steady-state baseband sequence. By absorbing the systematic disturbances introduced by hardware non-ideals at this stage, the generated time-frequency maps are ensured to mainly reflect the noise, interference, and target signal structure characteristics in the real electromagnetic environment, thereby improving the stability and repeatability of subsequent spectrum enhancement and detection.

[0011] Furthermore, step S102 specifically involves: employing a length of... The signal is divided into frames using a sliding time window, and a frame shift is set between adjacent frames. This achieves an engineering compromise between time resolution and frequency resolution. The frame signal is represented as: To reduce spectral leakage and improve the visibility of weak signals in the frequency domain, a window function is applied to each frame of the signal. Then perform a fast Fourier transform to obtain the corresponding complex spectrum representation: in This represents the frequency index. Designed for preprocessed steady-state baseband signals, it can effectively track their slowly changing spectral characteristics, providing a reliable time-frequency representation for subsequent feature extraction.

[0012] Furthermore, step S103 specifically involves: after obtaining the complex spectrum representation, performing an amplitude square operation on it to construct a power spectral density matrix. Because the energy range of different frequency bands and time segments is large in real wireless environments, in order to compress the dynamic range and improve the contrast of weak signal structures in the image domain, a logarithmic scaling of the power spectral density is performed to form a two-dimensional original time-frequency diagram: in To prevent the instability of small constants in logarithmic operations, their values ​​can be configured based on quantization precision and noise floor level. In some implementations, after generating the original time-frequency plot, the noise floor in the time or frequency dimension can be statistically estimated, and background subtraction or normalization can be performed to reduce the impact of noise floor differences between different monitoring scenarios on the subsequent model input distribution.

[0013] Furthermore, S201 specifically involves: using a high-quality time-frequency graph as a priori spectral structure during the training phase, assuming the reference time-frequency graph is... The reference time-frequency diagram can be obtained through high signal-to-noise ratio measured acquisition or simulation, and is used to characterize the spectral structure features under ideal conditions; to construct training input consistent with the actual monitoring environment, a spectral degradation operator is introduced. The reference time-frequency plot is degraded to generate a corresponding low-quality time-frequency plot: The parameter set It includes at least one or more of the following factors: noise intensity, interference superposition, quantization accuracy compression, dynamic range compression, and background fluctuation. The aforementioned reference time-frequency plot and low-quality time-frequency plot constitute the supervised learning data foundation for the augmentation network, providing conditions for the stable learning of subsequent augmentation mappings.

[0014] Furthermore, S301 specifically involves inputting the original time-frequency graph generated by signal acquisition and time-frequency transformation in S1 into the spectral feature enhancement model trained in S2, thereby obtaining an enhanced time-frequency graph as the basis for detection input. By establishing the detection process on enhanced spectral constraints, the detection model can focus more on physically meaningful signal structure regions in complex electromagnetic environments, rather than random noise or transient interference.

[0015] Furthermore, S302 specifically involves: under the condition of enhanced time-frequency map input, the target detection network performs regional-level analysis on the time-frequency plane and outputs several candidate signal regions; each candidate region is represented in the form of a bounding box, and its parameterized form is: in These represent the center positions of the candidate signal on the time axis and frequency axis, respectively. This indicates the coverage area of ​​the signal in both time and frequency directions; corresponding to each candidate region, the detection network simultaneously outputs the detection confidence level for that region. and signal category identifier The category identifier includes at least one or more of the following: UAV control signals, UAV image transmission signals, ordinary communication signals, and broadband interference signals; the target detection network can adopt a single-stage region detection architecture to achieve rapid discovery of multi-target signal regions, thereby obtaining the initial candidate signal set as follows: Used for subsequent result verification and output processing. Processing through the same pre-trained network framework unifies the processing logic and simplifies the design; at the same time, by replacing or retraining the network model, it can adapt to new signal types or more complex electromagnetic environments without rewriting the core algorithm, making the upgrade and maintenance path clear, and realizing the automated, scalable, and efficient conversion of information from low-level representation to high-level semantics.

[0016] Furthermore, step S303 specifically involves: filtering candidate regions based on a preset confidence threshold to remove detection results with insufficient confidence; then performing non-maximum suppression or an equivalent overlap suppression strategy on the remaining candidate regions to eliminate redundant detections generated in the same or highly overlapping time-frequency regions; finally, obtaining the confirmed set of signal targets. Each item corresponds to a time interval and frequency range. ), and the accompanying category determination results By adjusting the threshold, operators can flexibly balance between a high detection rate with no missed reports but potentially false alarms at a low threshold, and a high accuracy rate with reliable results but potentially missed reports at a high threshold, to adapt to different scenario requirements.

[0017] As can be seen from the above technical solution, the present invention has the following beneficial effects: 1. The present invention provides a method for detecting and recognizing UAV signals based on spectral feature enhancement. It adopts a learning-based spectral enhancement mechanism to improve the structural separability in low signal-to-noise ratio and spectral overlap scenarios from the source. Unlike traditional smoothing, noise reduction, and statistical filtering methods, the present invention uses a data-driven enhancement model to suppress noise, restore structural continuity, and unify contrast in degraded time-frequency maps. This transforms the originally blurry and fragmented UAV signals into regions with clear and locatable structures, fundamentally breaking through the recognition bottleneck of "unclear visibility and undetectable" under low signal-to-noise ratio conditions.

[0018] 2. This invention provides a UAV signal detection and recognition method based on spectrum feature enhancement, establishing a "structure-to-object" target-oriented output mechanism to enhance engineering usability and interoperability. It not only outputs signal categories but also directly generates structured target descriptions including time intervals, frequency ranges, categories, and confidence levels. The recognition results can be seamlessly integrated with upper-level systems such as situation analysis, alarm and countermeasure linkage, realizing the conversion from spectrum response to directly driven engineering targets and improving overall closed-loop processing efficiency.

[0019] 3. This invention provides a UAV signal detection and recognition method based on enhanced spectral features. This method reduces reliance on prior knowledge and fixed templates, improving adaptability to unknown and complex scenarios. It does not rely on frequency hopping sequence assumptions, protocol parsing, or RF fingerprint databases, but only on the enhanced physical layer time-frequency structure for detection. This allows the invention to maintain stable and reliable recognition performance even in complex scenarios such as encrypted signals, proprietary protocols, unknown UAV models, and multi-source concurrent interference, significantly reducing the requirements for database maintenance and feature completeness. Attached Figure Description

[0020] Figure 1 is a flowchart illustrating a method for detecting and recognizing unmanned aerial vehicle (UAV) signals based on spectral feature enhancement according to the present invention. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention. Embodiments

[0022] In this embodiment, as shown in Figure 1, the present invention discloses a method for UAV signal detection and recognition based on spectral feature enhancement, including: S1): signal acquisition and time-frequency transformation, specifically including: S101): preprocessing of the RF receiving link and baseband signal to obtain a steady-state baseband signal; S102): frame windowing and short-time Fourier transform to obtain a complex spectrum representation to characterize the joint distribution characteristics of the signal in the time and frequency dimensions; S103): power spectrum construction and generation of the original time-frequency map; S2): spectral feature enhancement, specifically including: S201): training sample construction, using a high-quality time-frequency map as a priori spectral structure, the reference time-frequency map is denoted as... ; Perform spectral degradation modeling to generate the corresponding degradation time-frequency graph; S202): Implement forward mapping through a spectral enhancement network structure, assuming the input degradation time-frequency graph is Let the spectrum enhancement network be denoted as... Its output enhancement time-frequency diagram is as follows: The spectrum enhancement network employs an encoder-decoder deep neural network structure.

[0023] Specifically, in S202, the encoding part extracts time-frequency features at different scales and compresses the spatial dimension through multi-layer convolution and downsampling operations. The decoding part gradually restores the resolution through upsampling and performs cross-layer fusion with the features of the corresponding encoding layer to preserve key details. To balance noise suppression and structure preservation, residual connections, multi-scale convolutions, and attention mechanisms can be configured within the network to suppress random noise and irrelevant background while highlighting the texture, bandwidth envelope, and possible frequency variation features of the UAV communication link in the spectrum. Through the above structural design, the enhanced network can complete the structural reconstruction of the degraded time-frequency map in a single forward inference process, meeting the requirements of real-time or near-real-time applications.

[0024] In this embodiment, S2 further includes S203): loss function design and enhancement target constraint, employing a combined optimization target composed of multiple losses. Specifically, pixel-level reconstruction loss is used to measure the difference between the enhanced time-frequency map and the reference time-frequency map at the level of overall energy distribution and amplitude scale, and is used to constrain the enhancement result to maintain consistency with the reference time-frequency map at the global energy level. This loss serves as a fundamental constraint term in the spectral feature enhancement process, preventing overall energy shift, amplitude imbalance, or abnormal amplification during the enhancement process.

[0025] In time-frequency plot representation, pixel-level differences reflect the energy deviation between the enhancement result and the reference result at the corresponding time-frequency position, and their statistical cumulative effect determines the overall energy distribution characteristics of the enhanced spectrum. By constraining these differences, it is possible to effectively ensure that the enhancement result can restore the structure and enhance the features of the degraded spectrum without changing the energy scale and statistical distribution of the original spectrum. In a typical implementation, one definition of the pixel-level reconstruction loss function is as follows: in, This represents the enhanced time-frequency diagram. This represents the reference time-frequency plot. By minimizing this loss term, the augmentation process can maintain consistency with the reference spectrum in terms of global energy and amplitude distribution, thereby providing a stable benchmark for subsequent augmentation targets based on structural similarity constraints and physical structure constraints.

[0026] The structural similarity loss is used to measure the similarity between the enhanced time-frequency map and the reference time-frequency map at the local statistical structure level, and to constrain the enhancement result to maintain consistency with the reference time-frequency map in general structural characteristics such as local brightness distribution, contrast relationship, and structural arrangement. The loss is not targeted at a specific communication system or spectral semantic structure, but rather serves as a general structural stability constraint to prevent local structural misalignment, morphological collapse, or abnormal structural responses during the enhancement process.

[0027] In time-frequency plot representation, local structural characteristics are mainly reflected in the mean energy distribution, contrast relationship, and correlation between energy changes at different locations within a finite time-frequency window. While these structural characteristics do not directly correspond to specific spectral physical semantics, they play a fundamental role in maintaining the overall readability, local morphological stability, and consistency with the reference time-frequency plot of the enhancement result. Based on these structural characteristics, the structural similarity loss imposes mesoscale structural constraints on the enhancement result by jointly measuring the brightness consistency, contrast relationship, and structural correlation between the enhanced and reference time-frequency plots within a local window. This maintains the stability of the enhanced spectrum at the local statistical structure level without introducing specific spectral priors, providing a reliable foundation for subsequent spectral structure enhancement and target detection. In a typical implementation, one definition of the structural similarity loss function is: in, This represents a measure of the structural similarity between the enhanced time-frequency image and the reference time-frequency image within a local window. By minimizing this loss term, the enhancement process can effectively suppress local structural distortion, morphological misalignment, and anomalous statistical responses introduced by the enhancement process, providing a stable mesoscale structural basis for the subsequent texture-aware loss based on physical structural constraints.

[0028] Texture-aware loss is used to measure the difference between the enhanced time-frequency map and the reference time-frequency map at the spectral structure level, and to constrain the enhancement result to maintain consistency with the reference time-frequency map in key structural features such as spectral line continuity, bandwidth envelope morphology, and fine-grained spectral texture. This loss focuses on spectral structural features closely related to UAV communication signal identification, but which are easily disrupted under low signal-to-noise ratio or strong interference conditions. In the time-frequency map representation, the aforementioned spectral structure has clear physical and geometric meanings: spectral line continuity reflects the continuous distribution of energy along the time or frequency axis; bandwidth envelope morphology corresponds to the signal's occupied range in the frequency dimension and its boundary variation trend; and fine-grained spectral texture reflects local energy fluctuations and periodic structures caused by modulation methods, symbol structures, or short-term transmission behavior. These structural features play a crucial role in subsequent signal region detection and category discrimination.

[0029] To explicitly constrain the structure recovery performance during training, this invention introduces feature extraction mapping. This is used to convert the input time-frequency graph into a set of structural response representations sensitive to changes in spectral structure. The structural responses characterize the geometric and morphological features in the time-frequency graph relevant to signal recognition, and their design aims to highlight spectral structural features such as energy gradient changes, directional consistency, cross-scale structural continuity, and band boundary responses. Based on these feature representations, feature extraction mapping... Constructed to be sensitive to local energy gradients, directional responses, and multi-scale structural changes, the mapping operator can be implemented through convolution operations, multi-scale filtering, directional response modeling, or combinations thereof, and can be implemented through neural network structures, learnable filters, or other equivalent signal processing modules. Regardless of the specific implementation, The purpose of all these methods is to transform the original pixel-level representation into a feature representation that has a clear response to changes in the spectral structure. During training, by enhancing the difference between the time-frequency image and the reference time-frequency image in the aforementioned structural response space, this invention utilizes a texture-aware loss function to constrain the enhancement results, thereby improving the separability and stability of the enhanced spectrum in subsequent detection and decision tasks. One definition of the texture-aware loss function is... By reducing the difference between the enhanced time-frequency map and the reference time-frequency map in the structural response space, the separability and stability of the enhanced spectrum in subsequent detection and decision tasks are significantly improved.

[0030] Based on the above loss description, this invention uses a comprehensive loss function to uniformly optimize the multi-layered constraint objectives in the spectral feature enhancement process, and its definition is as follows: in, The pixel-level reconstruction loss is used to constrain the enhancement results to maintain consistency with the reference time-frequency map in terms of overall energy distribution and amplitude scale. The structural similarity loss is used to constrain the stability of the enhancement results at the local statistical structure level. The texture-aware loss is used to constrain the enhancement results to maintain consistency with the reference time-frequency map at the level of spectral structure features with clear physical meaning. These are the corresponding weighting coefficients, used to balance global energy stability, mesoscale structural consistency, and high-level spectral structure discriminability.

[0031] In a typical implementation, the weighting coefficients can be set according to the following principles: Take a larger value to ensure the stability of the enhancement results in the overall energy and amplitude distribution; A moderate value is chosen to suppress local structural distortion; A relatively small value is chosen to enhance physically meaningful spectral textures and structural features without compromising overall energy and structural stability. For example, in one feasible embodiment, the weighting coefficient can be selected as follows: Or meet The proportional relationship.

[0032] In other implementations, the weighting coefficients can also be adaptively adjusted based on the signal-to-noise ratio (SNR) level of the training samples, the degree of spectral degradation, or the emphasis of the enhancement objective. For example, when the overall SNR of the training samples is low or the spectral degradation is severe, the weighting coefficients can be increased individually or simultaneously. and The value of is chosen to enhance the constraints on local structural stability and the recovery effect of spectral physical structure; when more attention is paid to overall energy stability or to avoid energy imbalance or abnormal structural response during the enhancement process, the value can be increased accordingly. The specific values ​​of the aforementioned weighting coefficients can be determined through empirical setting, cross-validation, or online adjustment; this invention does not limit their specific numerical form.

[0033] In this embodiment, the method further includes S3): signal region detection and target determination based on enhanced spectrum, specifically including: S301): target detection input and modeling based on enhanced spectrum constraints; S302): target region localization and generation of an initial candidate signal set; S303): candidate result confirmation and structured output.

[0034] In this embodiment, S101 specifically involves: the received radio frequency signal being sequentially subjected to pre-selected bandpass filtering to suppress strong out-of-band interference, low-noise amplification, and mixing down-conversion processing to obtain a complex baseband analog signal covering the target frequency band. The baseband signal is converted to a sampling rate by an analog-to-digital converter. Quantization bit width Sampling and quantization are performed to obtain a discrete-time complex sequence: in The quantization operator is represented; the selection of the sampling rate and quantization bit width is constrained by hardware cost and power consumption, and can be significantly lower than the configuration required for broadband continuous sampling. This is an engineering prerequisite for subsequently introducing a spectral structure enhancement mechanism; baseband preprocessing is performed on the discrete signal before time-frequency analysis to obtain a steady-state baseband sequence. .

[0035] Specifically, the preprocessing process includes, but is not limited to: estimating and eliminating the signal mean to suppress the DC component; normalizing or automatically controlling the signal amplitude to avoid dynamic range compression caused by strong interference signals; and correcting the I / Q amplitude-phase mismatch using a generalized linear compensation model.

[0036] In this embodiment, S102 specifically involves: using a length of... The signal is divided into frames using a sliding time window, and a frame shift is set between adjacent frames. This achieves an engineering compromise between time resolution and frequency resolution. The frame signal is represented as: To reduce spectral leakage and improve the visibility of weak signals in the frequency domain, a window function is applied to each frame of the signal. Then perform a fast Fourier transform to obtain the corresponding complex spectrum representation: in Indicates frequency index.

[0037] Specifically, the following can be selected in the engineering implementation: The transform length is used to increase the frequency sampling density, and its corresponding frequency resolution is... .

[0038] The above-mentioned framing and transformation parameters are set according to the following principles: under the given sampling rate constraint, the frequency resolution can characterize the typical bandwidth and spectral texture features of the UAV communication link, while reasonable frame shift settings ensure sufficient time sensitivity to burst transmissions, periodic control signals and possible frequency jump behaviors.

[0039] In this embodiment, S103 specifically involves: after obtaining the complex spectrum representation, performing an amplitude square operation on it to construct a power spectral density matrix. Because the energy range of different frequency bands and time segments is large in real wireless environments, in order to compress the dynamic range and improve the contrast of weak signal structures in the image domain, a logarithmic scaling of the power spectral density is performed to form a two-dimensional original time-frequency diagram: in To prevent the small constants from causing instability in logarithmic operations, their values ​​can be configured according to the quantization precision and noise floor level.

[0040] Specifically, in some implementations, after generating the original time-frequency map, statistical estimation of the noise floor in the time or frequency dimension can be performed, followed by background subtraction or normalization to reduce the impact of noise floor differences between different monitoring scenarios on the subsequent model input distribution. The original time-frequency map obtained through the above processing constitutes the direct input to the subsequent spectral structure separability enhancement module.

[0041] It should be noted that this invention does not limit the use of short-time Fourier transform for time-frequency analysis. Any method that can generate a two-dimensional time-frequency image representation under given sampling conditions, such as continuous wavelet transform or other equivalent time-frequency analysis methods, can be considered an equivalent implementation of the "original time-frequency image" in this invention.

[0042] In this embodiment, S201 specifically involves: using a high-quality time-frequency graph as a priori spectral structure during the training phase, assuming the reference time-frequency graph is... The reference time-frequency diagram can be obtained through high signal-to-noise ratio measured acquisition or simulation, and is used to characterize the spectral structure features under ideal conditions; to construct training input consistent with the actual monitoring environment, a spectral degradation operator is introduced. The reference time-frequency plot is degraded to generate a corresponding low-quality time-frequency plot: The parameter set It includes at least one or more of the following factors: noise intensity, interference superposition, quantization accuracy compression, dynamic range compression, and background fluctuation.

[0043] Specifically, by parameterizing the degradation process, the training samples can cover typical scenarios with different signal-to-noise ratios, different degrees of spectral overlap, and different hardware conditions. The aforementioned training samples... This forms the basis of supervised learning data for augmented networks, providing conditions for the stable learning of subsequent augmented mappings.

[0044] In this embodiment, S301 specifically involves inputting the original time-frequency graph generated by signal acquisition and time-frequency transformation in S1 into the spectrum feature enhancement model trained in S2, thereby obtaining an enhanced time-frequency graph as the basis for detection input.

[0045] Specifically, the enhanced time-frequency map has been restored in terms of overall energy distribution, spectral contrast, and structural continuity. Under this condition, target detection modeling no longer relies primarily on instantaneous energy mutations, but rather on the stable texture, bandwidth structure, and time-frequency occupancy pattern presented in the enhanced spectrum as the main criteria.

[0046] The object detection network uses a two-dimensional enhanced time-frequency map As input, spatial feature modeling and regional prediction are performed to characterize the joint distribution characteristics of wireless signals in the time and frequency dimensions. These characteristics include the extension of energy along the time and frequency directions, the occupied range in the frequency dimension, and its boundary changes, thereby forming candidate response regions with well-defined ranges on the time-frequency plane. By basing the detection process on enhanced spectral constraints, the detection model can focus more on physically meaningful signal structure regions in complex electromagnetic environments, rather than random noise or transient interference.

[0047] In this embodiment, S302 specifically involves: under the condition of enhanced time-frequency map input, the target detection network performs regional-level analysis on the time-frequency plane and outputs several candidate signal regions; each candidate region is represented in the form of a bounding box, and its parameterized form is: in These represent the center positions of the candidate signal on the time axis and frequency axis, respectively. This indicates the coverage area of ​​the signal in both time and frequency directions; corresponding to each candidate region, the detection network simultaneously outputs the detection confidence level for that region. and signal category identifier The category identifier includes at least one or more of the following: UAV control signals, UAV image transmission signals, ordinary communication signals, and broadband interference signals; the target detection network can adopt a single-stage region detection architecture to achieve rapid discovery of multi-target signal regions, thereby obtaining the initial candidate signal set as follows: Used for subsequent result confirmation and output processing.

[0048] Specifically, a two-stage detection method or a Transformer-based region modeling approach can be used to progressively screen and refine candidate regions. This invention does not limit the specific implementation of the detection network, but requires that its detection process take enhanced spectrum as input and output a set of candidate signals with clear time-frequency boundaries and category information.

[0049] In this embodiment, a confidence threshold can be set. Candidate regions are eliminated when their confidence level is below a certain threshold; for example... The value can be between 0.3 and 0.6. The specific value can be adjusted according to the application scenario and the tolerance for false detection. This invention does not limit the specific value.

[0050] After confidence screening, overlap suppression processing is performed on the remaining candidate regions to eliminate redundant detection results generated in the same or highly overlapping time-frequency regions. The degree of overlap can be measured by the intersection-over-union (IoU) ratio of the candidate regions in the time-frequency two-dimensional plane. When the IoU value between candidate regions is higher than a preset threshold... When the two candidate regions are considered to correspond to the same signal target instance, the overlapping candidate regions are merged into one instance for output. Specifically, the candidate region with the highest detection confidence in the group can be retained as the final output bounding box, and the category determination result corresponding to the candidate region can be used as the final category. When the category determination results of the group of candidate regions are inconsistent, the detection confidence is used as the sole criterion for decision, and the category with the highest confidence is selected as the final category, while the remaining candidate regions are suppressed.

[0051] In other implementations, the bounding box parameters of the same set of overlapping candidate regions can be weighted and fused or averaged to obtain a single output bounding box, and the final category can be determined using a confidence-based approach. All of these methods can be considered equivalent implementations. When the IoU value between candidate regions is lower than or equal to the threshold, they are considered to correspond to different signal target instances, and each candidate region is output as an independent instance.

[0052] After the confidence level screening and overlap suppression processes described above, the filtered set of signal targets is obtained: Each of these items corresponds to a specific time interval and frequency range. It outputs a stable category determination result. .

[0053] Specifically, the confirmation results are encapsulated into structured detection events for external output. The fields of these events include at least signal category, confidence level, time range, frequency interval, and duration. When a detection event exists in the confirmation results that belongs to the category of UAV-related signals and has a confidence level exceeding a preset threshold, a corresponding alarm is generated and reported to the upper-level monitoring platform, command system, or UAV countermeasures equipment via a communication interface, enabling subsequent coordinated action.

[0054] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting and recognizing UAV signals based on spectral feature enhancement, characterized in that: include: S1): Signal acquisition and time-frequency transformation, specifically including: S101): Preprocessing of the RF receiving link and baseband signal to obtain a steady-state baseband signal; S102): Framing, windowing, and short-time Fourier transform to obtain a complex spectrum representation to characterize the joint distribution characteristics of the signal in the time and frequency dimensions; S103): Power spectrum construction and generation of the original time-frequency plot; S2): Spectral feature enhancement, specifically including: S201): Construction of training samples, using a high-quality time-frequency plot as a priori spectral structure, with the reference time-frequency plot denoted as... ; Perform spectral degradation modeling to generate the corresponding degradation time-frequency graph; S202): Implement forward mapping through a spectral enhancement network structure, assuming the input degradation time-frequency graph is Let the spectrum enhancement network be denoted as... Its output enhancement time-frequency diagram is as follows: The spectrum enhancement network adopts an encoder-decoder deep neural network structure; S203): Loss function design and enhancement target constraints, adopting a combined optimization target composed of multiple losses, specifically: introducing pixel-level reconstruction loss to constrain the consistency of the enhanced time-frequency map and the reference time-frequency map in the overall energy distribution; introducing structural similarity loss to constrain spectral edges, spectral line continuity and local structural morphology; introducing texture perception loss, which measures the difference between the enhancement result and the reference result in the high-dimensional feature space through the feature extraction network, to emphasize the texture restoration effect related to signal recognition; the comprehensive loss function is defined as the weighted sum of the weight coefficients with the pixel-level reconstruction loss, structural similarity loss and texture perception loss respectively; S3): Signal region detection and target determination based on enhanced spectrum, specifically including: S301): Target detection input and modeling based on enhanced spectrum constraints; S302): Target region localization and generation of an initial candidate signal set; S303): Candidate result confirmation and structured output.

2. The method for detecting and recognizing UAV signals based on spectral feature enhancement according to claim 1, characterized in that: Specifically, S101 involves: the received radio frequency signal being sequentially subjected to a pre-selected bandpass filter to suppress strong out-of-band interference, low-noise amplification, and frequency mixing and down-conversion processing to obtain a complex baseband analog signal covering the target frequency band. The baseband signal is converted to a sampling rate by an analog-to-digital converter. Quantization bit width Sampling and quantization are performed to obtain a discrete-time complex sequence: in The quantization operator is used; baseband preprocessing is performed on the discrete signal before time-frequency analysis to obtain a steady-state baseband sequence. 。 3. The method for detecting and recognizing UAV signals based on spectral feature enhancement according to claim 1, characterized in that: Specifically, S102 involves using a length of... The signal is divided into frames using a sliding time window, and a frame shift is set between adjacent frames. This achieves an engineering compromise between time resolution and frequency resolution; The frame signal is represented as: To reduce spectral leakage and improve the visibility of weak signals in the frequency domain, a window function is applied to each frame of the signal. Then perform a fast Fourier transform to obtain the corresponding complex spectrum representation: in Indicates frequency index.

4. The method for detecting and recognizing UAV signals based on spectral feature enhancement according to claim 1, characterized in that: Specifically, S103 involves: after obtaining the complex spectrum representation, performing an amplitude square operation on it to construct the power spectral density matrix. Because the energy range of different frequency bands and time segments is large in real wireless environments, in order to compress the dynamic range and improve the contrast of weak signal structures in the image domain, a logarithmic scaling of the power spectral density is performed to form a two-dimensional original time-frequency diagram: in To prevent the small constants from causing instability in logarithmic operations, their values ​​can be configured according to the quantization precision and noise floor level.

5. The method for detecting and recognizing UAV signals based on spectral feature enhancement according to claim 1, characterized in that: Specifically, S201 involves: during the training phase, using a high-quality time-frequency graph as a priori spectral structure, assuming the reference time-frequency graph is... The reference time-frequency diagram can be obtained through high signal-to-noise ratio measured acquisition or simulation, and is used to characterize the spectral structure features under ideal conditions; to construct training input consistent with the actual monitoring environment, a spectral degradation operator is introduced. The reference time-frequency plot is degraded to generate a corresponding low-quality time-frequency plot: The parameter set It includes at least one or more of the following factors: noise intensity, interference superposition, quantization accuracy compression, dynamic range compression, and background fluctuation.

6. The method for detecting and recognizing UAV signals based on spectral feature enhancement according to claim 1, characterized in that: Specifically, S301 involves inputting the original time-frequency graph generated by signal acquisition and time-frequency transformation in S1 into the spectrum feature enhancement model trained in S2, thereby obtaining an enhanced time-frequency graph as the basis for detection input.

7. The method for detecting and recognizing UAV signals based on spectral feature enhancement according to claim 1, characterized in that: Specifically, S302 involves: under the condition of enhanced time-frequency map input, the target detection network performs regional-level analysis on the time-frequency plane and outputs several candidate signal regions; each candidate region is represented in the form of a bounding box, and its parameterized form is as follows: in These represent the center positions of the candidate signal on the time axis and frequency axis, respectively. This indicates the coverage area of ​​the signal in both time and frequency directions; corresponding to each candidate region, the detection network simultaneously outputs the detection confidence level for that region. and signal category identifier The category identifier includes at least one or more of the following: UAV control signals, UAV image transmission signals, ordinary communication signals, and broadband interference signals; the target detection network can adopt a single-stage region detection architecture to achieve rapid discovery of multi-target signal regions, thereby obtaining the initial candidate signal set as follows: Used for subsequent result confirmation and output processing.

8. The method for detecting and recognizing UAV signals based on spectral feature enhancement according to claim 1, characterized in that: S303 specifically involves: filtering candidate regions based on a preset confidence threshold to remove detection results with insufficient confidence; then applying non-maximum suppression or an equivalent overlap suppression strategy to the remaining candidate regions to eliminate redundant detections in the same or highly overlapping time-frequency regions; finally, obtaining the confirmed set of signal targets. Each item corresponds to a time interval and frequency range. ), and output the category determination result. 。

Images