Wide-spectrum rapid scanning and intelligent identification method, system and device

Abstract

The invention discloses a wide-spectrum rapid scanning and intelligent identification method, system and device, and the method comprises the steps: firstly, carrying out the rapid frequency sweeping through employing a low-sampling-rate and low-quantization-precision receiver, generating a low-resolution time-frequency graph, and recovering the super-resolution of the low-resolution time-frequency graph into a high-resolution spectrum through a pre-trained deep learning model; then, the enhanced time-frequency image is segmented into time-frequency blocks, global features are extracted through a Transform encoder, and detection and screening of multiple candidate signal areas are achieved; and finally, carrying out directional sampling on the candidate region, respectively extracting time domain and frequency domain features, mapping the time domain and frequency domain features to a unified space for fusion, and completing signal category judgment based on the combined features and a judgment network. According to the wide-frequency-spectrum rapid scanning and intelligent identification method, system and device, through a layered architecture of broadband rapid scanning-frequency spectrum quality recovery-multi-region detection-directional sampling and time-frequency fusion identification, the reliability of unmanned aerial vehicle signal detection and identification in a complex electromagnetic environment is improved while the hardware cost is reduced.

Description

A wide-spectrum fast scanning and intelligent identification method, system and device Technical Field

[0001] This invention belongs to the field of radio monitoring and early warning technology for unmanned aerial vehicles (UAVs), and specifically relates to a wide-spectrum rapid scanning and intelligent identification method, system, and device. 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, commercial 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, and oil and petrochemical parks has become increasingly prominent. This includes activities such as cross-border aerial photography, privacy theft, delivery of contraband into monitored areas, and detection and evasion of security systems, posing new challenges to public safety and the protection of key infrastructure. Illegal drones typically possess characteristics such as flexible take-off and landing, short flight time, and strong anonymity, putting significant pressure on traditional security systems.

[0003] To address the problem of unauthorized drone intrusion, counter-UAS (counter-drone) technology is constantly evolving, 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 detecting 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, offer advantages such as non-active radiation, wide coverage, and all-weather operation, making them a crucial technological direction for counter-UAS systems.

[0004] However, existing spectrum detection schemes are mostly focused on narrowband or fixed-frequency monitoring, which usually requires prior knowledge of the frequency band or communication system used by the drone. In practical applications, illegal drones may operate in ISM bands such as 2.4GHz and 5.8GHz, or they may use manufacturer-defined frequencies or frequency hopping systems. Their control and image transmission links are often distributed across multiple subbands, and they may even evade detection through wideband hopping strategies. This places higher demands on the system to perform rapid scanning and signal discovery over a wide frequency range. To achieve wideband real-time monitoring, traditional schemes often rely on broadband receivers with high sampling rates and high quantization accuracy, as well as high-speed analog-to-digital converters, to directly sample hundreds of MHz or even wider bandwidths. While such schemes can obtain high-quality spectrum data, they are costly and power-intensive, making them unsuitable for large-scale deployment. Some systems use frequency sweeping for subband polling sampling to reduce hardware requirements, but the spectrum generated under low-precision sampling conditions is prone to problems such as low resolution, high noise, and inconsistencies between subbands. Traditional methods based on energy thresholds or simple features are not robust enough in environments with multiple signals coexisting and low signal-to-noise ratios. Furthermore, in real-world electromagnetic environments, the 2.4GHz and 5.8GHz frequency bands are typically home to numerous Wi-Fi, Bluetooth, and other wireless devices. Their signals highly overlap with drone communication in the frequency domain and exhibit bursty and irregular characteristics in time. The control and image transmission signals of illegal drones typically have low power, resulting in a low signal-to-noise ratio at long distances or under strong interference conditions. This makes it difficult for detection methods based on spectrum templates or fixed bandwidth to balance false positives and false negatives.

[0005] In recent years, deep learning methods have been introduced into the fields of spectrum analysis and signal recognition, enabling signal classification through automatic feature extraction from time-frequency maps. However, existing methods largely rely on clear time-frequency maps acquired by high-precision receivers, resulting in significant performance degradation under low resolution and low signal-to-noise ratio conditions. Furthermore, most methods primarily classify single regions, lacking the ability to simultaneously detect and locate multiple signal regions across a wide spectrum. Moreover, illegal drone communication links typically exhibit significant time-domain and frequency-domain coupling characteristics, making reliable differentiation difficult when relying solely on features from a single domain.

[0006] In summary, existing UAV spectrum monitoring technologies generally suffer from high costs associated with wideband monitoring, spectrum quality degradation under low-precision frequency sweeping conditions, insufficient robustness in multi-source interference environments, and a lack of systematic solutions for wideband multi-region detection and joint time-domain and frequency-domain identification. Therefore, it is necessary to propose a UAV monitoring technology solution that achieves wideband rapid scanning, spectrum quality recovery, multi-region detection, and refined identification under low-cost hardware conditions. Summary of the Invention

[0007] Purpose of the invention: In order to overcome the above shortcomings, the purpose of this invention is to provide a wide-spectrum fast scanning and intelligent recognition method, system and device. Through the hierarchical architecture of "wide-spectrum fast scanning - spectrum quality recovery - multi-region detection - directional sampling and time-frequency fusion recognition", the reliability of UAV signal detection and recognition in complex electromagnetic environments is improved while reducing hardware costs.

[0008] Technical Solution: To achieve the above objectives, this invention provides a wide-spectrum fast scanning and intelligent identification method, comprising: S1): wide-spectrum fast frequency scanning and spectrum reconstruction, specifically including the following steps: S101): data acquisition and preprocessing, converting the continuous-time wide-spectrum wireless signal into a two-dimensional time-frequency diagram representation; S102): deep model training based on high-resolution spectrum data; in the model training stage, a high-performance receiver is used to perform high-resolution sampling of the target wide-spectrum band; assuming that high sampling rate and high quantization accuracy are respectively... and This yields a high-resolution discrete signal: Based on the description in the previous step, for Generate a high-resolution time-frequency plot, denoted as To construct low-resolution training samples corresponding to high-resolution samples, a controllable degradation operator is introduced. ,right Degradation processing is performed, including but not limited to operations such as reducing the sampling rate, compressing the quantization bit width, subbanding sampling, and introducing noise, to obtain: And generate the corresponding low-resolution time-frequency graph, denoted as By constructing training sample pairs For deep learning models Training is performed to learn the mapping relationship from low-resolution spectrum to high-resolution spectrum; S103): Fast frequency sweeping with large bandwidth under low-resolution conditions, using a receiver architecture with low sampling rate and low quantization accuracy to quickly sweep the target frequency band; the target frequency band is divided into multiple sub-bands, where the bandwidth of each sub-band is matched with the sampling rate; by controlling the local oscillator to quickly switch between the center frequencies of different sub-bands, and performing short-time sampling in each sub-band, the corresponding sub-band time-frequency map is generated; by stitching the sub-band time-frequency maps on the frequency axis, a low-resolution sweep frequency-frequency map covering the entire target frequency band is formed; S1 04): Super-resolution recovery of swept spectrum based on deep model, inputting the swept time-frequency map described in S103 into the trained deep model, and outputting an enhanced wide-spectrum time-frequency map; S2): Wide-spectrum multi-region detection, specifically including the following steps: S201): Wide-spectrum time-frequency map segmentation and feature extraction; the enhanced wide-spectrum time-frequency map is segmented into time-frequency blocks that meet the constraints; the constraints are: the frequency dimension size is not less than the sub-band bandwidth, and the time dimension size is not less than the dwell window, so as to preserve the complete spectrum structure and time continuity; the time-frequency blocks are flattened, embedded, and have two-dimensional position coding added to form The token sequence is input into a multi-layer Transformer encoder; through multi-head self-attention global modeling, the time-frequency correlation features of cross-subband energy concentration regions and frequency hopping signals are captured; S202): candidate signal region detection; S203): candidate signal region screening; S3): time-domain combined with frequency-domain fusion recognition, specifically including the following steps: S301): directional sampling and local signal acquisition based on candidate regions; S302): parallel extraction of time-domain and frequency-domain features; S303): fusion of time-domain and frequency-domain features and signal decision; after completing the extraction of time-domain and frequency-domain features, the two types of features are mapped to a unified feature space and feature fusion processing is performed to form a joint feature representation; the joint feature representation is input to the decision network to output the signal category determination result and its confidence information of the corresponding candidate region; in the signal decision process, the candidate region is classified according to the maximum a posteriori probability criterion; when the output probability corresponding to a certain signal category exceeds the preset decision threshold, the candidate region is determined to be the corresponding category signal; when the output probability of each category is lower than the threshold, the candidate region is determined to be a normal communication signal or background interference signal.

[0009] Furthermore, S101 specifically involves: receiving broadband radio signals containing interference from drones and other sources in scenarios such as airports and prisons; down-converting the broadband radio signals to baseband via an antenna, filter, amplifier, and other RF front-ends to obtain complex signals; subsequently sampling and quantizing the complex signals to obtain discrete sequences, and calculating their time-frequency energy distribution using short-time Fourier transform; finally, converting the continuous signal into a two-dimensional time-frequency graph, serving as a unified data format for subsequent model processing and analysis. S101 defines a standardized processing flow from raw RF signals to a two-dimensional time-frequency graph, providing a consistent and well-organized data foundation for all subsequent processing (deep learning models, detection, and recognition) by converting continuous and complex wireless signals into a unified time-frequency graph.

[0010] Furthermore, S202 specifically involves: based on the global feature representation output by the Transformer, configuring a region detection head to perform region-level prediction on the Token or combinations of adjacent Tokens, generating a set of candidate signal regions; for each candidate region, the detection head outputs the following set of parameters: in, This indicates the confidence level that the region is a valid signal area; This indicates the signal attribute identifier, used to distinguish between candidate UAV control signals, candidate UAV image transmission signals, ordinary communication signals, or interference signals; This indicates the start and end positions of the signal in the time dimension; This indicates the start and end positions of the signal in the frequency dimension; the set of candidate signal regions is: .

[0011] The detector head can not only determine the presence or absence of a signal, but also directly output a set of structured parameters, including confidence level, signal attribute label, and precise time-frequency boundaries. This provides direct and clear control instructions and prior information for subsequent directional sampling and classification, greatly improving the automation and efficiency of the entire processing chain.

[0012] Further, step S203 specifically involves: performing region filtering on the candidate signal region set; the filtering process includes: preliminary filtering based on a confidence threshold; a suppression strategy based on time-frequency overlap, retaining overlapping regions with higher confidence; and obtaining the final candidate signal region set after filtering, denoted as... This serves as the output of the wide-spectrum multi-region detection module. It effectively eliminates unreliable detection results and redundant boxes, ensuring that the final output candidate region set has higher accuracy and representativeness.

[0013] Furthermore, S301 specifically involves: for each candidate signal region, based on its frequency range... With time range Generate corresponding directional sampling control parameters; during real-time monitoring or the next scan, adjust the receiver's local oscillator center frequency to... And set the receiving bandwidth to cover the frequency span of that region. Within the corresponding time window, a higher sampling rate than that used in the wideband sweep phase is employed. Sampling is performed, where: This yields the original baseband signal segment corresponding to the candidate region, denoted as... The signal segments are used for time-domain feature modeling and generating local fine-grained time-frequency representations, serving as the input basis for subsequent time-frequency fusion recognition. This achieves on-demand allocation and optimized use of system resources, obtaining high-quality data of key signals without significantly increasing overall hardware costs and continuous data volume, which is crucial for balancing coverage and recognition accuracy.

[0014] Furthermore, S302 specifically involves: for the signal segments obtained from directional sampling, constructing time-domain feature branches and frequency-domain feature branches respectively to achieve parallel extraction of signal behavior characteristics and spectral structure characteristics; in the time-domain feature branch, inputting the original I / Q signal sequence or its amplitude and phase sequences into a one-dimensional sequence modeling network to extract the signal's behavioral features in the time dimension; in the frequency-domain feature branch, generating a local fine-grained time-frequency map based on the signal segments; the local fine-grained time-frequency map has a higher frequency and time resolution than the broadband sweep stage, used to characterize the signal's structural features in the spectral dimension. Time-domain features can better characterize the signal's modulation details, instantaneous behavior, and other "dynamic" characteristics; frequency-domain features are better at revealing the signal's spectral structure, energy distribution, and other "static" characteristics; the parallel dual-branch design provides a richer and more robust feature set for subsequent fusion recognition, helping to cope with signal variations and interference in complex environments.

[0015] This invention also provides a wide-spectrum fast scanning and intelligent recognition system for implementing the aforementioned wide-spectrum fast scanning and intelligent recognition method. The system includes: a radio signal receiving and RF front-end module, a wide-spectrum fast frequency sweep acquisition module, a wide-spectrum multi-region detection module, and a feature fusion and signal decision module connected sequentially. The radio signal receiving and RF front-end module implements step S101; the wide-spectrum fast frequency sweep acquisition module implements steps S102 to S104; the wide-spectrum multi-region detection module implements step S3; and the feature fusion and signal decision module implements step S4. The modular structure allows the system to be split or deployed as a whole according to actual needs such as edge computing or cloud processing, enhancing adaptability.

[0016] The present invention also provides an apparatus for implementing a wide-spectrum fast scanning and intelligent recognition method, the apparatus comprising: a processor configured to execute computer-executable instructions; and a memory storing one or more computer-executable instructions, wherein when the computer-executable instructions are executed by the processor, the various steps of the wide-spectrum fast scanning and intelligent recognition method are implemented.

[0017] As can be seen from the above technical solution, the present invention has the following beneficial effects: 1. The present invention provides a wide-spectrum fast scanning and intelligent recognition method, system and device, which adopts local oscillator fast frequency scanning and low-resolution sampling to achieve wide-spectrum coverage, significantly reducing hardware cost and power consumption; and through a super-resolution model based on deep learning, the low-quality frequency scanning spectrum is restored to a high-resolution time-frequency map, effectively overcoming the problems of sub-band splicing traces and weak signal loss, and providing a high-quality input foundation for subsequent processing.

[0018] 2. The present invention provides a wide-spectrum fast scanning and intelligent identification method, system and device. Based on the enhanced wide-spectrum time-frequency map, it uses the Transformer architecture for global modeling, which can detect and locate multiple time-frequency regions simultaneously in one processing. It can effectively adapt to complex electromagnetic scenarios such as coexistence of multiple signals, spectrum overlap and frequency hopping, and improve the comprehensiveness and robustness of detection.

[0019] 3. The present invention provides a wide-spectrum fast scanning and intelligent recognition method, system and device, which performs directional resampling of candidate regions under detection guidance, extracts the temporal behavior features and frequency domain structure features of the signal in parallel, and performs deep fusion and joint decision-making, which surpasses the limitations of a single feature domain and significantly improves the classification accuracy and anti-interference capability of key signals such as UAV control and image transmission.

[0020] 4. The present invention provides a wide-spectrum fast scanning and intelligent identification method, system and device, which mainly identifies based on the physical layer time and frequency characteristics, does not rely on specific protocol parsing, and has good model expansion and unknown signal adaptability; its structured output (such as region, category, confidence level) is easy to integrate with existing monitoring, alarm and countermeasure systems to form a complete monitoring, identification and handling closed loop. Attached Figure Description

[0021] Figure 1 is a schematic diagram of the steps of a wide-spectrum fast scanning and intelligent recognition method according to the present invention; Figure 2 is a schematic diagram of the architecture of a wide-spectrum fast scanning and intelligent recognition system according to the present invention. Detailed Implementation

[0022] 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

[0023] In this embodiment, as shown in Figure 1, the present invention discloses a wide-spectrum fast scanning and intelligent identification method, including: S1): wide-spectrum fast frequency scanning and spectrum reconstruction, specifically including the following steps: S101): data acquisition and preprocessing, converting the continuous-time wide-spectrum wireless signal into a two-dimensional time-frequency diagram representation; S102): deep model training based on high-resolution spectrum data; in the model training stage, a high-performance receiver is used to perform high-resolution sampling of the target wide-spectrum band; let the high sampling rate and high quantization accuracy be respectively... and This yields a high-resolution discrete signal: Based on the description in the previous step, for Generate a high-resolution time-frequency plot, denoted as To construct low-resolution training samples corresponding to high-resolution samples, a controllable degradation operator is introduced. ,right Degradation processing is performed, including but not limited to operations such as reducing the sampling rate, compressing the quantization bit width, subbanding sampling, and introducing noise, to obtain: And generate the corresponding low-resolution time-frequency graph, denoted as By constructing training sample pairs For deep learning models The model is trained to learn the mapping relationship from low-resolution to high-resolution spectra. The deep learning model is used to establish this mapping relationship; its network structure is not limited and can be implemented using an encoder-decoder convolutional neural network, a residual network, or a hybrid network structure incorporating an attention mechanism. The model takes a low-resolution time-frequency plot as input and the corresponding high-resolution time-frequency plot as output, and training is completed in an end-to-end manner.

[0024] In one training method, the training data comes from real radio spectrum data collected within the target monitoring area, and low-resolution samples are constructed through the aforementioned degradation operation. During training, a composite loss function, consisting of pixel reconstruction error and spectral structure consistency constraints, is used to ensure that the reconstructed spectrum is consistent with the high-resolution reference spectrum in terms of energy distribution and structural continuity. The optimization algorithm can employ stochastic gradient descent or the Adam optimizer, and the learning rate is adaptively adjusted based on the training convergence. Model training ends when the validation error no longer decreases significantly.

[0025] The deep learning model trained in the above manner can effectively recover spectral structure details under low-resolution frequency sweep conditions and improve the discriminability of weak signals, providing high-quality spectral input for subsequent wide-spectrum multi-region detection and fine recognition.

[0026] In this embodiment, S1 further includes S103): fast frequency sweeping with large bandwidth under low resolution conditions, using a receiving architecture with low sampling rate and low quantization accuracy to perform fast frequency sweeping of the target frequency band; dividing the target frequency band into multiple sub-bands, as shown in the following formula: in, Indicates the target frequency band to be monitored. This represents the k-th sweep sub-band, where K is the total number of sub-bands. The bandwidth of each sub-band is related to the low sampling rate. This ensures that signals within a single sub-band can be effectively acquired and corresponding time-frequency representations generated under the current sampling conditions.

[0027] By controlling the local oscillator to rapidly switch between different sub-band center frequencies and maintaining a preset dwell time within each sub-band for short-term sampling and time-frequency analysis, a corresponding sub-band time-frequency map is generated. Subsequently, the time-frequency maps of each sub-band are stitched together according to their order on the frequency axis to form a low-resolution swept-frequency time-frequency map covering the entire target frequency band, denoted as... ,in Indicates a time-dimensional index. This represents a frequency-dimensional index.

[0028] During the aforementioned frequency sweeping process, due to the constraints of wideband coverage requirements and the simultaneous use of low sampling rate, low quantization accuracy, and sub-band polling sampling mechanisms, the frequency sweep time-frequency map typically suffers from some degradation in spectral quality. Specifically, this manifests as insufficient frequency resolution, high noise levels, and amplitude discontinuities between adjacent sub-bands. This degradation makes it difficult for the low-resolution frequency sweep time-frequency map to directly support the accurate detection and identification of weak or complex signal structures. Therefore, it is necessary to introduce spectral reconstruction and super-resolution processing in subsequent steps to improve structural continuity and the usability of the sweeping results.

[0029] In this embodiment, S1 further includes S104): super-resolution recovery of the swept frequency spectrum based on the deep model, by inputting the swept frequency time-frequency map described in S103 into the trained deep model: The meanings of the symbols in the above formula are explained below: Indicates low sampling rate and low quantization accuracy Under these conditions, a low-resolution wideband time-frequency map is obtained by sub-band polling frequency sweeping, where... For time dimension index, Indexed by frequency dimension; The enhanced broadband time-frequency map obtained after spectrum reconstruction and enhancement processing has the same coverage in the time and frequency dimensions as the low-resolution swept frequency time-frequency map, but it is improved in terms of spectrum structure continuity, amplitude consistency and weak signal discernibility. This represents a deep learning-based spectrum reconstruction model used to establish a mapping relationship between low-resolution swept-frequency time-frequency maps and high-quality spectrum representations; the model is trained offline.

[0030] This represents the set of model parameters for the deep learning model, which is obtained during the model training phase through offline training using high-resolution spectral data and corresponding low-resolution degraded samples, and remains fixed during the practical application phase. It is used to perform spectral reconstruction and enhancement processing on the low-resolution time-frequency map obtained during the frequency sweep phase.

[0031] Through the above mapping relationship, the system reconstructs the spectral structure in the low-resolution sweep frequency map without changing the wideband fast frequency sweep acquisition method, so as to restore and enhance its structural continuity, weak signal discernibility and cross-subband consistency, thereby providing a stable and reliable input for the subsequent wideband multi-region detection and UAV signal intelligent recognition module.

[0032] In this embodiment, the method further includes S2): wide-spectrum multi-region detection, specifically including the following steps: S201): wide-spectrum time-frequency map segmentation and feature extraction; firstly, the enhanced wide-spectrum time-frequency map is processed... Two-dimensional segmentation along the time and frequency axes yields several time-frequency blocks: The size of each time-frequency block is... The size of this time-frequency block is directly related to the sweep frequency parameters in S3, and its constraints are as follows: (1) Frequency dimension size Not less than the bandwidth of a single swept subband, to ensure that the time-frequency block contains the spectral structure within the complete subband; (2) Time dimension size The minimum time-frequency window is not less than the dwell time of the first sub-band, so as to ensure that the time continuity characteristics of the control signal or frequency hopping signal can be characterized.

[0033] Next, each time-frequency block is flattened and mapped to a unified dimension via linear embedding. The token representation is obtained as follows: in, This is the embedding matrix.

[0034] To explicitly preserve time-frequency location information, a two-dimensional time-frequency location code is introduced for each token. The final token is represented as: This forms the token sequence: The token sequence is input into a multi-layer Transformer encoder, which models the correlation between different time-frequency blocks globally through a multi-head self-attention mechanism. Through this global feature modeling process, the model can simultaneously capture the following wide-spectrum structural features: (1) energy concentration regions distributed across multiple sub-bands; (2) the correlation between frequency-hopping signals in the time and frequency dimensions; and (3) structural differences under the condition of multiple signal coexistence.

[0035] Through the above methods, the transition from "enhanced broadband perception" to "targeted fine recognition" is completed, forming a processing link with broadband scanning as the starting point and regional-level analysis as the core, thereby improving the accuracy and efficiency of UAV signal detection in complex environments with multiple signals coexisting and overlapping spectra.

[0036] In this embodiment, S2 further includes: S202): candidate signal region detection; S203): candidate signal region filtering.

[0037] In this embodiment, the method further includes S3): time-domain combined with frequency-domain fusion recognition, specifically including the following steps: S301): directional sampling and local signal acquisition based on candidate regions; S302): parallel extraction of time-domain and frequency-domain features; S303): fusion of time-domain and frequency-domain features and signal decision; after completing the extraction of time-domain and frequency-domain features, the two types of features are mapped to a unified feature space, and feature fusion processing is performed to form a joint feature representation; the joint feature representation is input to the decision network to output the signal category determination result and its confidence information for the corresponding candidate region; the specific structure of the decision network is not limited, and it can be implemented using a classification structure such as a multi-layer fully connected neural network, the input of which is the fused joint feature vector, and the output is the probability value corresponding to each type of signal; in the signal decision process, the candidate region is classified according to the maximum a posteriori probability criterion; when the output probability corresponding to a certain signal category exceeds a preset decision threshold, the candidate region is determined to be the corresponding category signal; when the output probability of each category is lower than the threshold, the candidate region is determined to be a normal communication signal or a background interference signal.

[0038] Through the above time-domain-frequency domain feature fusion and decision mechanism, the following signal types can be distinguished and identified: (1) illegal UAV control link signal; (2) illegal UAV image transmission link signal; (3) ordinary communication signal or background interference signal.

[0039] The aforementioned time-frequency domain fusion identification mechanism can improve the accuracy and stability of UAV signal identification in complex electromagnetic environments where spectrum overlap is severe and single-domain features are difficult to reliably distinguish, and provide a reliable basis for decision-making for subsequent early warning or response modules.

[0040] In this embodiment, S101 specifically involves: firstly receiving and preprocessing radio environment signals within the target monitoring area. These radio environment signals include actual drone communication signals present in scenarios such as airports, prisons, power facilities, oil and petrochemical parks, and key urban areas, as well as other wireless communications and background interference signals partially overlapping with their spectrum. The radio frequency front-end includes a multi-band antenna, a pre-selection filter, a low-noise amplifier (LNA), a switchable subband filter, a local oscillator, and a mixer, used to receive target frequency band signals and perform down-conversion processing. The receiving antenna is positioned within the target monitoring frequency band. The internal received radio frequency signal is The signal first passes through a bandpass filter to suppress out-of-band interference in the RF front end, and then is amplified by a low-noise amplifier to obtain the amplified signal. Subsequently, using the local oscillator signal right The target frequency band is down-converted to baseband to obtain a complex baseband signal. For baseband signals at sampling rate Quantization accuracy Perform analog-to-digital conversion to obtain a discrete-time signal sequence: Based on this, a Short-Time Fourier Transform (STFT) or equivalent time-frequency analysis operation is performed on the discrete signal to generate the corresponding time-frequency representation: in, The window function is used; through the above processing, the continuous-time broadband wireless signal is converted into a two-dimensional time-frequency graph representation, which serves as a unified data form for subsequent model training, broadband frequency sweeping, and spectrum reconstruction.

[0041] In this embodiment, S202 specifically involves: configuring a region detection head based on the global feature representation output by the Transformer, performing region-level prediction on the Token or combinations of adjacent Tokens, and generating a set of candidate signal regions; for each candidate region, the detection head outputs the following set of parameters: Where, represents the confidence level that the area is a valid signal area; represents the signal attribute identifier, used to distinguish candidate UAV control signals, candidate UAV image transmission signals, ordinary communication signals or interference signals; This indicates the start and end positions of the signal in the time dimension; This indicates the start and end positions of the signal in the frequency dimension; the set of candidate signal regions is: .

[0042] In this embodiment, S203 specifically involves performing region filtering on the candidate signal region set. The region filtering includes the following steps: (1) Preliminary filtering based on a confidence threshold: For each candidate region in the candidate signal region set, the region-level confidence is compared with a preset confidence threshold, and candidate regions with confidence levels lower than the threshold are eliminated. The confidence threshold is used to distinguish between potential effective signal regions and noise or false detection regions. Its specific value can be set based on the statistical results of the model training and verification phase, and is used to balance the candidate region recall rate and false detection rate. The confidence threshold can also be configured or adjusted according to the interference intensity, background noise level, or system operation strategy in the actual electromagnetic environment. For example, in scenarios with complex electromagnetic environments and strong background interference, the confidence threshold can be appropriately increased to reduce false detections caused by interference signals. In scenarios where it is necessary to improve the detection capability of weak signals, the confidence threshold can be appropriately decreased to improve the coverage capability of potential UAV signals.

[0043] (2) Suppression processing based on time-frequency overlap: For candidate signal regions that have passed the initial filtering by the confidence threshold, the system analyzes the coverage of different candidate regions in the time and frequency dimensions. Specifically, for any two candidate regions, if their corresponding time intervals overlap and their corresponding frequency intervals also overlap, they are considered to have an overlapping relationship in the time-frequency domain. In the case of multiple time-frequency overlapping candidate regions, the system processes the candidate regions in descending order of confidence, prioritizing the retention of candidate regions with higher confidence, and suppressing or eliminating the remaining candidate regions that have an overlapping relationship in the time-frequency domain to reduce redundant output.

[0044] Through the above time-frequency overlap suppression processing, the system can effectively reduce repeated outputs caused by detection redundancy or boundary offset while maintaining the ability to detect effective signal regions, thereby improving the simplicity and reliability of the candidate signal region set.

[0045] The final set of candidate signal regions obtained after screening is denoted as . It serves as the output of the wide-spectrum multi-region detection module. It not only indicates the location of potential signals in the enhanced wide spectrum, but also provides clear control interface parameters for subsequent analysis, including: (1) the frequency range that needs to be analyzed in detail; (2) the corresponding time window; and (3) the region-level confidence level and signal identification.

[0046] Based on the above output, the receiver can be guided to perform directional sampling and fine identification within the corresponding frequency band and time window in subsequent processing, thereby forming a processing link of "wideband scanning - area discovery - directional fine analysis".

[0047] In this embodiment, S301 specifically refers to: for each candidate signal region According to its frequency range With time range The corresponding directional sampling control parameters are generated. The generation of these directional sampling control parameters is based on the time-frequency boundary information of the candidate signal region. Specifically, the frequency range of the candidate signal region... As a constraint, it is mapped to the receiver's local oscillator center frequency and receiving bandwidth control parameters, where the local oscillator center frequency is set to the center position of the frequency range to ensure that the candidate signal region is centrally covered within the receiving bandwidth: At the same time, the receiving bandwidth is set to cover the frequency range. To fully encompass the spectral span of the candidate signal region.

[0048] Further, based on the time range of the candidate signal region As a time window constraint for directional sampling, the receiver is triggered to perform a sampling operation within this time window; the sampling operation uses a sampling rate higher than that of the wideband sweep phase. To obtain local signal data with higher time and frequency resolution, where: In this way, the time-frequency boundary information of the candidate signal region is directly converted into the receiver's local oscillator control, bandwidth control, and sampling timing control parameters, thereby achieving directional sampling and local signal acquisition of the candidate signal region. The original baseband signal segment corresponding to the candidate region obtained in this way is denoted as... The signal segment is used for time-domain feature modeling and generating a local fine time-frequency representation, which serves as the input basis for the subsequent time-frequency fusion recognition module.

[0049] In this embodiment, step S302 specifically involves: constructing a time-domain feature branch and a frequency-domain feature branch for the signal segment obtained by directional sampling, respectively, to achieve parallel extraction of signal behavior characteristics and spectral structure characteristics; in the time-domain feature branch, the original I / Q signal sequence or its amplitude and phase sequences are input into a one-dimensional sequence modeling network to extract the signal's behavior characteristics in the time dimension; in the frequency-domain feature branch, a local fine time-frequency map is generated based on the signal segment; the local fine time-frequency map has a higher frequency and time resolution than the wideband sweep stage, and is used to characterize the signal's structural features in the spectral dimension.

[0050] Specifically, in the time-domain feature branch, the extracted features are deep features, and the signal features represented include, but are not limited to: (1) the periodic or quasi-periodic features of the control signal; (2) the continuous high load and burst interval features of the image transmission signal; (3) the asymmetry of uplink and downlink data traffic; and (4) the burst, idle and switching rhythm of the signal.

[0051] The specific network structure of the temporal feature extraction network is not limited and can be implemented using a one-dimensional convolutional neural network, a temporal convolutional network (TCN), or other sequence modeling structures. The network is trained based on samples of UAV control signals, image transmission signals, and ordinary communication signals collected under different monitoring scenarios, and the network parameters are optimized through supervised learning, thereby improving the ability to characterize the temporal behavior differences of different signal types.

[0052] Specifically, the frequency domain feature extraction network encodes the local time-frequency map and extracts features including but not limited to the following: (1) local spectral texture and energy distribution pattern; (2) subcarrier structure and bandwidth occupancy features; (3) frequency hopping or frequency drift mode; and (4) spectral stability features that are different from common communication signals.

[0053] The specific structure of the frequency domain feature extraction network is not limited and can be implemented using a two-dimensional convolutional neural network or a lightweight Transformer structure. The network is trained based on time-frequency map samples obtained from actual frequency sweeping and directional sampling, and supervised learning is completed by combining the corresponding signal type labels to enhance the ability to represent the differences in the frequency domain structure of different signals in complex spectral environments.

[0054] In this embodiment, as shown in Figure 2, the present invention also discloses a wide-spectrum fast scanning and intelligent recognition system for the aforementioned wide-spectrum fast scanning and intelligent recognition method. The system includes: a radio signal receiving and RF front-end module, a wide-spectrum fast frequency sweep acquisition module, a wide-spectrum multi-region detection module, and a feature fusion and signal decision module connected in sequence; the radio signal receiving and RF front-end module is used to implement S101; the wide-spectrum fast frequency sweep acquisition module is used to implement S102 to S104; the wide-spectrum multi-region detection module is used to implement S3; and the feature fusion and signal decision module is used to implement S4.

[0055] Specifically, the system first receives wireless signals within the target frequency band through a radio signal receiving and RF front-end module, and performs down-conversion processing. Subsequently, the signal enters a wideband fast sweep frequency acquisition module, which generates a low-resolution wideband time-frequency map under low sampling rate conditions. The enhanced wideband representation is then output by a spectrum enhancement and reconstruction module.

[0056] Based on this, the system simultaneously detects multiple candidate signal regions in the enhanced wide spectrum using a wide-spectrum multi-region detection module, and outputs the corresponding time window, frequency range, and confidence information. The candidate region results are used to filter valid signal regions and to drive the receiver to perform directional sampling.

[0057] For signals obtained by directional sampling, the system performs time-domain feature extraction and frequency-domain feature extraction in parallel, and completes the fusion decision in the feature fusion and signal decision module, and finally outputs the identification results of UAV control link signal, UAV image transmission link signal or ordinary communication / background interference signal.

[0058] In this embodiment, the present invention also discloses an apparatus for implementing a wide-spectrum fast scanning and intelligent recognition method. The apparatus includes: a processor configured to execute computer-executable instructions; and a memory storing one or more computer-executable instructions, wherein when the computer-executable instructions are executed by the processor, the various steps of the wide-spectrum fast scanning and intelligent recognition method are implemented.

[0059] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements can be made 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 wide-spectrum fast scanning and intelligent recognition method, characterized in that: include: S1): Wideband fast frequency sweep and spectrum reconstruction, specifically including the following steps: S101): Data acquisition and preprocessing, converting the continuous-time wideband wireless signal into a two-dimensional time-frequency diagram representation; S102): Deep model training based on high-resolution spectrum data; in the model training stage, a high-performance receiver is used to perform high-resolution sampling of the target wideband; let the high sampling rate and high quantization accuracy be respectively... and This yields a high-resolution discrete signal: Based on the description in the previous step, for Generate a high-resolution time-frequency plot, denoted as To construct low-resolution training samples corresponding to high-resolution samples, a controllable degradation operator is introduced. ,right Degradation processing is performed, including but not limited to operations such as reducing the sampling rate, compressing the quantization bit width, subbanding sampling, and introducing noise, to obtain: And generate the corresponding low-resolution time-frequency graph, denoted as By constructing training sample pairs For deep learning models Training is performed to learn the mapping relationship from low-resolution spectrum to high-resolution spectrum; S103): Fast frequency sweeping with large bandwidth under low-resolution conditions, using a receiver architecture with low sampling rate and low quantization accuracy to quickly sweep the target frequency band; the target frequency band is divided into multiple sub-bands, where the bandwidth of each sub-band is matched with the sampling rate; by controlling the local oscillator to quickly switch between the center frequencies of different sub-bands, and performing short-time sampling in each sub-band, the corresponding sub-band time-frequency map is generated; by stitching the sub-band time-frequency maps on the frequency axis, a low-resolution sweep frequency-frequency map covering the entire target frequency band is formed; S1 04): Super-resolution recovery of swept spectrum based on deep model, inputting the swept time-frequency map described in S103 into the trained deep model, and outputting an enhanced wide-spectrum time-frequency map; S2): Wide-spectrum multi-region detection, specifically including the following steps: S201): Wide-spectrum time-frequency map segmentation and feature extraction; the enhanced wide-spectrum time-frequency map is segmented into time-frequency blocks that meet the constraints; the constraints are: the frequency dimension size is not less than the sub-band bandwidth, and the time dimension size is not less than the dwell window, so as to preserve the complete spectrum structure and time continuity; the time-frequency blocks are flattened, embedded, and have two-dimensional position coding added to form The token sequence is input into a multi-layer Transformer encoder; through multi-head self-attention global modeling, the time-frequency correlation features of cross-subband energy concentration regions and frequency hopping signals are captured; S202): candidate signal region detection; S203): candidate signal region screening; S3): time-domain combined with frequency-domain fusion recognition, specifically including the following steps: S301): directional sampling and local signal acquisition based on candidate regions; S302): parallel extraction of time-domain and frequency-domain features; S303): fusion of time-domain and frequency-domain features and signal decision; after completing the extraction of time-domain and frequency-domain features, the two types of features are mapped to a unified feature space and feature fusion processing is performed to form a joint feature representation; the joint feature representation is input to the decision network to output the signal category determination result and its confidence information of the corresponding candidate region; in the signal decision process, the candidate region is classified according to the maximum a posteriori probability criterion; when the output probability corresponding to a certain signal category exceeds the preset decision threshold, the candidate region is determined to be the corresponding category signal; when the output probability of each category is lower than the threshold, the candidate region is determined to be a normal communication signal or background interference signal.

2. The wide-spectrum fast scanning and intelligent recognition method according to claim 1, characterized in that: Specifically, S101 involves receiving broadband radio signals containing interference from drones and other sources in scenarios such as airports and prisons; down-converting the broadband radio signals to baseband via an antenna, filter, amplifier, and other RF front-end to obtain complex signals; subsequently sampling and quantizing the complex signals to obtain discrete sequences, and calculating their time-frequency energy distribution through short-time Fourier transform; finally, the continuous signal is converted into a two-dimensional time-frequency graph, serving as a unified data format for subsequent model processing and analysis.

3. The wide-spectrum fast scanning and intelligent recognition method according to claim 1, characterized in that: Specifically, S202 involves configuring a region detection head based on the global feature representation output by the Transformer, performing region-level prediction on the Token or combinations of adjacent Tokens, and generating a set of candidate signal regions. For each candidate region, the detection head outputs the following set of parameters: in, This indicates the confidence level that the region is a valid signal area; This indicates the signal attribute identifier, used to distinguish between candidate UAV control signals, candidate UAV image transmission signals, ordinary communication signals, or interference signals; This indicates the start and end positions of the signal in the time dimension; This indicates the start and end positions of the signal in the frequency dimension; the set of candidate signal regions is: 。 4. The wide-spectrum fast scanning and intelligent recognition method according to claim 1, characterized in that: S203 specifically involves: performing region filtering on the candidate signal region set; the filtering process includes: preliminary filtering based on a confidence threshold; a suppression strategy based on time-frequency overlap, retaining overlapping regions with higher confidence; and obtaining the final candidate signal region set after filtering, denoted as... It serves as the output of the wide-spectrum multi-region detection module.

5. The wide-spectrum fast scanning and intelligent recognition method according to claim 1, characterized in that: Specifically, S301 involves: for each candidate signal region, based on its frequency range... With time range Generate corresponding directional sampling control parameters; during real-time monitoring or the next scan, adjust the receiver's local oscillator center frequency to... ； And set the receive bandwidth to cover the frequency span of that region. Within the corresponding time window, a higher sampling rate than that used in the wideband sweep phase is employed. Sampling is performed, where: This yields the original baseband signal segment corresponding to the candidate region, denoted as... The signal segment is used for time-domain feature modeling and generating a local fine-grained time-frequency representation, which serves as the input basis for subsequent time-frequency fusion recognition.

6. The wide-spectrum fast scanning and intelligent recognition method according to claim 5, characterized in that: Specifically, S302 involves: constructing a time-domain feature branch and a frequency-domain feature branch for the signal segment obtained by directional sampling, thereby realizing the parallel extraction of signal behavior characteristics and spectral structure characteristics; in the time-domain feature branch, the original I / Q signal sequence or its amplitude and phase sequences are input into a one-dimensional sequence modeling network to extract the signal's behavior characteristics in the time dimension. In the frequency domain feature branch, a local fine time-frequency map is generated based on the signal segment; the local fine time-frequency map has a higher frequency and time resolution than the wideband sweep stage, and is used to characterize the structural features of the signal in the spectral dimension.

7. A wide-spectrum fast scanning and intelligent recognition system, used to implement the wide-spectrum fast scanning and intelligent recognition method according to any one of claims 1 to 6, characterized in that: The system includes: a radio signal receiving and RF front-end module, a wideband fast frequency sweep acquisition module, a wideband multi-region detection module, and a feature fusion and signal decision module connected in sequence; the radio signal receiving and RF front-end module is used to implement S101; the wideband fast frequency sweep acquisition module is used to implement S102 to S104; the wideband multi-region detection module is used to implement S3; and the feature fusion and signal decision module is used to implement S4.

8. An apparatus for implementing a wide-spectrum fast scanning and intelligent recognition method, characterized in that: The apparatus includes: a processor configured to execute computer-executable instructions; and a memory storing one or more computer-executable instructions, which, when executed by the processor, implement the various steps of the wide-spectrum fast scanning and intelligent recognition method according to any one of claims 1 to 6.

Images