Unmanned aerial vehicle detection method based on multi-band radio spectrum analysis

Abstract

The invention discloses an unmanned aerial vehicle detection method based on multi-band radio spectrum analysis, and the method can recognize the flight control behaviors of a target, such as hovering, cruising or hovering, through the extraction of a signal track and the introduction of a behavior modeling mechanism, and effectively overcomes the defect that the prior art lacks the dynamic behavior analysis capability. A multi-node synchronization mechanism and a TDOA positioning method are combined, so that high-precision traceability of a target space position is realized; meanwhile, risk level assessment is carried out based on signal features, a track mode and a space region, so that the system has intelligent early warning capability. According to the whole scheme, a full-link unmanned aerial vehicle detection method from multi-frequency receiving, frequency spectrum identification, behavior judgment, target positioning to risk output is constructed, and the problems of single frequency band, rough identification means, lack of tracking and evaluation capability and the like in the prior art are effectively solved; and the practicability and the accuracy of the unmanned aerial vehicle detection system in a complex airspace environment are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) detection technology, and in particular to a UAV detection method based on multi-band radio spectrum analysis. Background Technology

[0002] In recent years, with the gradual advancement of low-altitude airspace opening policies and the widespread development of civilian drone technology, an increasing number of small multi-rotor drones have been widely used in logistics delivery, power line inspection, film and television shooting, environmental monitoring, and other fields. To maintain airspace order and ensure public safety, it is urgent to deploy drone detection systems with real-time capabilities, high coverage, and high accuracy to proactively perceive and intelligently identify drone behavior in the airspace.

[0003] Currently, mainstream drone detection methods mainly include radar echo recognition, video image recognition, acoustic array localization, and radio signal analysis. Among these, radar technology has weak reflectivity against targets with small flight attitudes and low speeds, limiting its recognition range; video recognition is limited by factors such as lighting, weather, and obstruction, and its real-time processing is poor; acoustic detection lacks sensitivity in scenarios with high flight altitudes or strong background noise. In contrast, radio spectrum analysis technologies offer higher concealment and deployment flexibility, enabling the detection of drone remote control links or image transmission links without actively emitting signals, and are gradually becoming the main research direction for current anti-black flight detection systems. However, existing radio detection solutions generally suffer from the following problems: First, most systems only cover fixed frequency bands, such as 2.4GHz or 5.8GHz, and cannot effectively identify illegal drones using low frequency bands such as 433MHz or 915MHz, or even custom frequency hopping mechanisms; Second, most spectrum identification algorithms use simple energy detection or rule-based feature matching, lacking the ability to model complex modulation, multi-frequency synchronization, or encrypted communication behaviors, and are easily affected by background interference; Third, the system as a whole lacks comprehensive capabilities such as multi-node collaborative perception, dynamic target tracking, intelligent behavior discrimination, and risk level assessment, making it difficult to support the practical deployment needs of urban or strategic airspace.

[0004] Therefore, in UAV radio detection systems, the complexity of dynamic changes in spectrum signals, the concealment of multi-band communication behavior, the difficulty in identifying non-standard protocols, and the lack of distributed linkage analysis and risk level assessment capabilities have become urgent problems to be solved. Summary of the Invention

[0005] This application provides a method for detecting unmanned aerial vehicles (UAVs) based on multi-band radio spectrum analysis.

[0006] A method for UAV detection based on multi-band radio spectrum analysis, the method comprising:

[0007] S1, multiple spectrum receiving nodes are arranged in a region to be monitored, wherein the multiple nodes have multi-band radio acquisition capability for acquiring multiple band signal data including 2.4 GHz, 5.8 GHz, 433 MHz, 915 MHz, 1.2 GHz in parallel;

[0008] S2, the collected signal of each frequency band is subjected to noise reduction, normalization and frame window segmentation processing to generate a multi-channel time-frequency spectrum graph in a unified format;

[0009] S3, the multi-channel time-frequency spectrum graph is input into a spectrum fusion analysis model containing convolution structure and Transformer structure, local features within the frequency band and time sequence features between the frequency bands are extracted respectively, and target discrimination results containing confidence and labels are output;

[0010] S4, for a target discriminated as a suspected unmanned aerial vehicle control signal or a picture transmission signal, target carrier frequency, symbol rate, frequency offset trajectory and modulation mode signal features are extracted, and a time sequence signal trajectory is constructed;

[0011] S5, the target trajectory is sent to a behavior modeling module, the target spectrum behavior is modeled by a state transition analysis method, and a suspected flight control behavior is identified, the control behavior including hovering, cruising and returning modes;

[0012] S6, based on the time stamps of the same target signal received by the multiple spectrum receiving nodes, a TDOA algorithm is used to calculate the two-dimensional or three-dimensional positioning information of the signal source, and the target motion path is reconstructed in combination with historical trajectory data;

[0013] S7, a risk level is calculated for the identified target signal, and an unmanned aerial vehicle detection event record containing target spatial position, signal category, trajectory feature and threat level is generated.

[0014] In the above scheme, optionally, the radio frequency front end of the spectrum receiving node comprises:

[0015] A multi-channel parallel radio frequency receiving link, each link is configured with a low noise amplifier, a variable bandwidth bandpass filter, an automatic gain control unit, a programmable local oscillator frequency source and an analog-to-digital conversion module;

[0016] The bandpass filter supports on-demand switching of center frequency and bandwidth parameters for dynamically adapting different frequency band signals;

[0017] The local oscillator frequency source and the analog-to-digital conversion module are both provided with a unified reference clock by a high-precision temperature-compensated crystal oscillator;

[0018] Data sampling synchronization and time stamp injection are realized between the multiple links through an FPGA logic controller, and docking with GPS, Beidou PPS pulse or IEEE 1588 time protocol is supported;

[0019] The collected spectrum data is packaged in an IQ pair format, with a channel number, a frequency band identifier, and a high-precision timestamp field, for subsequent central analysis and processing.

[0020] In the above scheme, the spectrum fusion analysis model comprises:

[0021] at least one parallel convolution branch for different frequency band time-frequency spectrum, for extracting local texture features and edge features of the spectrum;

[0022] a Transformer encoder module with a multi-head self-attention structure, for capturing temporal evolution rules and cross-frequency band correlations;

[0023] a classification and discrimination head, for outputting multi-class label results including legal communication, UAV remote control, image transmission channel, and unknown signals.

[0024] In the above scheme, the behavior modeling module comprises:

[0025] a set of state sequence generation units, for converting signal trajectories into state vector sequences;

[0026] an LSTM-based behavior classification network, for predicting the behavior type corresponding to the target according to the state sequence;

[0027] The behavior type includes sustained hovering, low-speed cruising, variable-speed return, or short-time interference burst mode.

[0028] In the above scheme, the TDOA positioning module comprises:

[0029] a distributed node time synchronization mechanism, supporting the use of multiple synchronization protocols, including but not limited to GPS PPS pulse, IEEE 1588 PTP protocol, and local crystal oscillator taming mechanism;

[0030] Each receiving node has an independent time verification module, which can dynamically calibrate the sampling time deviation;

[0031] The nodes periodically broadcast synchronization status, spectrum sensing status, and network health information, for judging the current synchronization quality;

[0032] The TDOA positioning algorithm module receives the same signal timestamps, spatial positions, and signal strengths uploaded by at least three nodes, and estimates the signal source coordinates using a weighted positioning algorithm;

[0033] The weighted algorithm adjusts the contribution weight according to the synchronization error estimate value, signal-to-noise ratio level, and distance of each node;

[0034] The positioning result is provided with a credibility level field, and a track point can be output for path backtracking or track splicing analysis.

[0035] In the foregoing solution, optionally, the method further includes:

[0036] S8, a spectrum behavior fingerprint database is constructed to record carrier variation characteristics, modulation modes, frequency hopping modes, and symbol rate interval attributes of historical unmanned aerial vehicle communication signals.

[0037] S9, the spectrum characteristics of the current identified target signal are matched with the fingerprint database, if the matching is successful, the target signal is marked as a known manufacturer model or protocol family, and if the matching is not successful, the target signal is marked as a new type.

[0038] In the foregoing solution, optionally, the calculation of the risk level is based on:

[0039] Similarity of the current target signal and a historical legal list;

[0040] Duration of existence of the signal and frequency activity level;

[0041] Whether a space region covered by a signal track contains a no-fly zone or a warning line;

[0042] Whether a high-risk behavior mode is identified, specifically including: multi-band synchronous image transmission or frequency band hopping encryption behavior.

[0043] In the foregoing solution, optionally, the detection event record includes a behavior credibility score field for evaluating reliability of target signal identification, and the score is generated by the following sub-modules in cooperation:

[0044] A model discrimination confidence module outputs a neural network softmax or sigmoid result of an identification label as a basic score;

[0045] A signal track continuity evaluation module calculates a continuity score based on factors such as frequency track smoothness, modulation mode stability, and duration length;

[0046] A space consistency judgment module performs trend consistency matching on a target cross-node positioning path;

[0047] A behavior model fitting degree module evaluates whether a target time sequence feature can be highly fitted and predicted by an existing behavior model;

[0048] The score results of the dimensions are calculated into a total credibility score according to preset weights, and are attached with a confidence interval, which is used as a reference for event disposal decision.

[0049] In the foregoing solution, optionally, the plurality of spectrum receiving nodes construct a distributed spectrum cooperative sensing system, including:

[0050] Each receiving node is provided with a local spectrum preliminary screening module for tagging the identified target signal with a confidence label and a positioning timestamp;

[0051] The identification result is uploaded to the central node through a message queue or a lightweight publish-subscribe protocol, supporting breakpoint resuming;

[0052] The central node constructs an identification graph structure with nodes as vertices and frequency similarity and time sequence proximity as edges;

[0053] Based on a graph neural network, the identification results in the graph are fused and optimized to generate a unified detection label and a consistency score;

[0054] The central node can reversely schedule the front-end node to resample or adjust the frequency band monitoring parameters, realizing adaptive resource allocation.

[0055] In the above scheme, the method further comprises:

[0056] S10, the trajectories of the targets identified by different receiving nodes in a plurality of geographical areas are spliced and analyzed, if the signal trajectories have frequency band consistency, frequency hopping feature similarity and spatial continuity, it is determined that they are cross-regional detection results of the same unmanned aerial vehicle target, and are recorded.

[0057] Compared with the prior art, the present application has at least the following beneficial effects:

[0058] Based on further analysis and research of the problems of the prior art, it is realized that in the unmanned aerial vehicle radio detection system, the dynamic change complexity of the spectrum signal, the concealment of the multi-frequency communication behavior, the difficulty of non-standard protocol identification, and the lack of distributed linkage analysis and risk level discrimination ability, by constructing a multi-frequency spectrum receiving architecture, combining a spectrum fusion analysis model with a convolution and Transformer hybrid structure, the precise identification of unmanned aerial vehicle remote control signals and image transmission signals under multi-frequency and multi-modulation conditions is realized, solving the problems of insufficient frequency band coverage and low identification accuracy of traditional methods; by extracting signal trajectories and introducing a behavior modeling mechanism, the flight control behavior of the target, such as hovering, cruising or returning, can be identified, effectively supplementing the lack of dynamic behavior analysis capability of the prior art; combined with a multi-node synchronization mechanism and a TDOA positioning method, high-precision tracing of the spatial position of the target is realized; at the same time, risk level assessment is carried out based on signal features, trajectory patterns and spatial areas, so that the system has intelligent early warning capability. The overall scheme constructs a full-link unmanned aerial vehicle detection method from multi-frequency reception, spectrum identification, behavior judgment, target positioning to risk output, effectively addressing the problems of single frequency, rough identification means and lack of tracking and evaluation capability in the background technology, significantly improving the practicality and accuracy of the unmanned aerial vehicle detection system in complex airspace environments. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 A flowchart of a method for detecting a UAV based on multi-band radio spectrum analysis is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0060] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application.

[0061] In the description of the present application: unless otherwise specified, the meaning of "multiple" is two or more. The terms "first", "second", "third", etc. in the present application are intended to distinguish the objects referred to, and do not have a special meaning in the technical connotation aspect (for example, it should not be understood as emphasizing importance or order, etc.). The expressions "include", "contain", "have" and the like also mean "not limited to" (some units, components, materials, steps, etc.).

[0062] In one embodiment, as shown in Figure 1 a method for detecting a UAV based on multi-band radio spectrum analysis is provided, comprising the following steps:

[0063] S1, deploying a plurality of spectrum receiving nodes in a region to be monitored, wherein the plurality of nodes have multi-band radio acquisition capability for parallel acquisition of multi-band signal data including 2.4GHz, 5.8GHz, 433MHz, 915MHz, 1.2GHz, etc.

[0064] S2, performing noise reduction, normalization and frame window segmentation processing on the collected signal of each frequency band to generate a multi-channel time-frequency spectrum graph in a unified format;

[0065] S3, inputting the multi-channel time-frequency spectrum graph into a spectrum fusion analysis model containing convolution structure and Transformer structure, respectively extracting local features within the frequency band and time sequence features between the frequency bands, and outputting target discrimination results containing confidence and labels;

[0066] S4, for targets discriminated as suspected UAV control signals or image transmission signals, extracting target carrier frequency, symbol rate, frequency offset trajectory and modulation mode signal features, and constructing a time series signal trajectory;

[0067] S5, sending the target trajectory into a behavior modeling module, modeling the target spectrum behavior through a state transition analysis method, for identifying its suspected flight control behavior, the control behavior including hovering, cruising and homing modes;

[0068] S6, based on the timestamps of the same target signal received by multiple spectrum receiving nodes, using TDOA algorithm to calculate the two-dimensional or three-dimensional positioning information of the signal source, and combining historical trajectory data to realize target motion path reconstruction;

[0069] S7, calculating the risk level of the identified target signal, and generating a UAV detection event record containing target spatial position, signal category, trajectory characteristics and threat level.

[0070] The embodiment proposes a UAV detection method based on multi-band radio spectrum analysis, aiming at the defects of traditional detection methods in frequency band coverage, identification accuracy, target continuity tracking and behavior understanding, and constructs a complete detection process covering "data collection-spectrum identification-behavior modeling-target positioning-risk assessment".

[0071] First, in step S1, multiple spectrum receiving nodes are deployed in the airspace area to be monitored (such as urban core area, airport perimeter, event venue, etc.). Each node is internally integrated with a radio frequency module supporting multi-band reception, which can cover 2.4GHz, 5.8GHz, 433MHz, 915MHz, 1.2GHz, etc. Common or commonly used control signal frequency bands. The receiving link in each node supports a multi-channel parallel receiving architecture, each channel selects the target frequency band through a band-pass filter, enhances the weak signal receiving capability through an automatic gain control (AGC) module, uses a phase-locked local oscillator and an analog-to-digital converter to complete the sampling from radio frequency to baseband, and the sampling result is output in the form of IQ pair with timestamp for subsequent multi-node collaborative analysis. All nodes are synchronized through PPS pulse or IEEE 1588 protocol to achieve high-precision synchronization, laying a foundation for subsequent positioning and trajectory splicing.

[0072] In step S2, the original signals collected by each channel are subjected to a unified data preprocessing process. First, the signal is denoised (such as wavelet denoising), and then amplitude normalization and standardization operations are performed to keep the signal amplitudes of different channels consistent. Then, frame operation is performed according to a fixed time window (e.g. 100ms), and short-time Fourier transform (STFT) or wavelet transform method is applied to construct two-dimensional time-frequency spectrum, each spectrum reflects the frequency distribution of a frequency band signal in a period of time. The time-frequency spectrum corresponding to multiple frequency bands is collected, i.e. a multi-channel time-frequency spectrum input sequence is formed for subsequent deep model analysis.

[0073] In step S3, the multi-channel time-frequency spectrogram is input into a spectrum fusion analysis model composed of multiple parallel convolution extraction branches combined with a Transformer structure. Specifically, a convolution network is used to extract local texture features, jump structure, energy distribution patterns, and other static features in the spectrogram; the Transformer structure is used to model the time sequence change relationship between different frequency bands, including frequency synchronization, frequency hopping interval, amplitude change trend, and other dynamic behaviors. The overall model adopts a multi-head attention mechanism to improve cross-frequency learning capability, and outputs the spectrum type label and its confidence in each time window. The classification results include "legal communication signal", "unmanned vehicle remote control link", "image backhaul link", and "unknown signal".

[0074] When the model identifies a suspected remote control or image transmission type illegal target, it enters step S4, and models the signal of the target in a continuous time sequence to extract its carrier frequency, symbol rate, modulation mode, frequency offset trajectory, and other features in multiple time windows. These features are combined to form a "spectrum trajectory sequence", which is processed by the behavior modeling module.

[0075] In step S5, the behavior modeling module is used to analyze whether the spectrum change behavior of the target presents a typical flight control intention. For example, an LSTM network or a state transition graph reasoning model can be used for sequence classification of the spectrum trajectory to determine whether it has a structured control process such as "takeoff-cruise-hover-return". If a typical behavior chain is identified, it can further assist in determining whether it is a known unmanned vehicle brand or has a special task behavior pattern.

[0076] In step S6, the identified target is located. Based on the target signal timestamp information and node coordinate information uploaded by at least three receiving nodes, a two-dimensional or three-dimensional position is estimated using a time difference of arrival (TDOA) algorithm. Each node locally marks the first arrival time of the target, and the central server analyzes the time difference of the node data and constructs a hyper-surface intersection calculation model to inversely deduce the position of the target signal source. To improve positioning accuracy, the TDOA module will perform weighted calculation based on node signal strength, synchronization error estimation, and other indicators, and output a reliability index and a positioning error estimation.

[0077] Finally, in step S7, the system calculates the comprehensive risk level based on the spectrum label, behavior characteristics, spatial position, and trajectory characteristics of the target. The risk assessment model can use a scoring and weighting method, and the reference factors include whether the target is in a sensitive area, whether its modulation mode is abnormal, whether the behavior has a task pattern, whether there is a multi-band cooperative communication behavior, etc. Finally, a detection event containing the target position, signal category, judgment confidence, risk level, and other information is output for use by the platform warning module or to link with emergency response equipment.

[0078] Through the coordinated efforts of the above steps, this method can achieve spectrum perception, behavior determination, and risk assessment of complex, covert, and non-standard protocol UAVs. It has the technical advantages of high recognition rate, high coverage, and high interpretability, and is particularly suitable for airspace supervision and intelligent countermeasure systems in scenarios such as urban low-altitude airspace, key infrastructure protection areas, and military no-fly zones.

[0079] In this embodiment, the radio frequency front-end of the spectrum receiving node includes:

[0080] A multi-channel parallel RF receiving link, each link is equipped with a low-noise amplifier, a variable bandwidth bandpass filter, an automatic gain control unit, a programmable local oscillator frequency source, and an analog-to-digital converter module;

[0081] The bandpass filter supports on-demand switching of center frequency and bandwidth parameters to dynamically adapt to signals in different frequency bands.

[0082] Both the local oscillator frequency source and the analog-to-digital conversion module are provided with a unified reference clock by a high-precision temperature-compensated crystal oscillator.

[0083] Data sampling synchronization and timestamp injection between multiple links are achieved through FPGA logic controllers, and it supports interfacing with GPS, BeiDou PPS pulse or IEEE 1588 time protocol;

[0084] The acquired spectrum data is encapsulated in IQ pair format and includes channel number, frequency band identifier, and high-precision timestamp fields for subsequent central analysis and processing.

[0085] In one specific implementation, the radio frequency front-end of each spectrum receiving node adopts a multi-channel parallel architecture, meaning that the node simultaneously has multiple independent radio frequency receiving links, each configured to target one or a group of frequency bands. For example, channel 1 can be fixed to collect the 2.4GHz frequency band, channel 2 can collect the 5.8GHz frequency band, and channel 3 supports dynamic switching between the 433MHz and 915MHz frequency bands to accommodate the remote control / image transmission communication characteristics of different brands of drones.

[0086] Each channel is equipped with a low-noise amplifier (LNA) at the front end to enhance weak signal reception and improve the system's minimum detectable signal level. A subsequently configured variable bandwidth bandpass filter supports programmable control of bandwidth and center frequency, isolating effective signals within specific frequency bands and suppressing adjacent out-of-band interference.

[0087] To adapt to the dramatic changes in signal strength in different environments, an automatic gain control module (AGC) is integrated in each signal link to dynamically adjust the gain parameters by monitoring the signal amplitude in real time, so that the sampling signal remains within the optimal dynamic range, avoiding oversaturation or distortion. In addition, the receiving link also integrates a local oscillator frequency control module, i.e. a phase-locked loop (PLL) + local oscillator signal source combination, whose center frequency can be dynamically adjusted through remote configuration instructions, working together with the filter to complete the frequency band switching capability.

[0088] The analog-to-digital conversion module (ADC) converts the processed analog signal into a digital IQ pair data stream, with each frame of data containing channel number, frequency band label, and timestamp information, which is generated by a unified reference clock. The data streams of all channels are aggregated to the node controller and uploaded to the central processing platform.

[0089] Synchronization between multiple channels is controlled by the FPGA logic control unit, which uses a high-precision temperature-compensated crystal oscillator (such as a 10 MHz TCXO) as a unified reference clock source to ensure that all channels complete synchronized acquisition within the same time sampling window. It also supports accessing GPS, Beidou PPS pulses, or IEEE 1588 PTP protocols to time-conform external clock sources, further ensuring time alignment of cross-node data.

[0090] This embodiment realizes high-sensitivity, concurrent, and stable acquisition of different frequency band signals in complex electromagnetic environments, while providing a unified time base, channel identification, and raw spectrum quality guarantee for signal classification and recognition, target trajectory construction, and TDOA positioning, significantly enhancing the overall perception capability of the system front end.

[0091] In this embodiment, the spectrum fusion analysis model includes:

[0092] At least one parallel convolution branch for different frequency band time-frequency spectrograms for extracting local texture features and edge features of the spectrogram;

[0093] A Transformer encoder module with a multi-head self-attention structure for capturing temporal evolution rules and cross-frequency correlations;

[0094] A classification and discrimination head for outputting multi-class label results including legal communication, UAV remote control, image transmission channel, and unknown signals.

[0095] The spectrum fusion analysis model consists of three main parts: a multi-channel parallel convolution branch structure, a multi-head self-attention Transformer encoding module, and a signal classification and discrimination head, which are combined in the order of "input—fusion—discrimination" process.

[0096] Firstly, the data from multiple frequency bands (such as 2.4GHz, 5.8GHz, 915MHz, etc.) are sampled through radio frequency links and processed into standardized time-frequency spectrograms, and input into the model in the form of multi-channel input. For each spectrogram corresponding to each frequency band, the system sets up a dedicated convolutional neural network branch (CNN Branch). Each CNN branch usually includes several convolutional layers, batch normalization layers, ReLU activation layers, and pooling layers, etc., for extracting local texture features, frequency jump edges, energy bandwidth changes, and other structured signal features in the spectrogram of the frequency band.

[0097] Because different frequency bands have different signal structure characteristics, for example, the remote control modulation method in the 433MHz frequency band is very different from the 5.8GHz image transmission method in the frequency domain, the multi-branch structure can retain the unique characteristics of each frequency band without being weakened by fusion, thereby forming a multi-frequency band local feature vector set.

[0098] After extracting the basic features of each frequency band, the feature vectors are sent to a unified multi-head self-attention mechanism (Multi-head Attention) Transformer encoder module. The design of this module draws on the advantages of the Transformer structure in natural language processing, and its core capability is to model the dependency relationship between any positions in the input sequence. In this invention, the self-attention mechanism is used to model the coupling relationship between different frequency band signals in time, such as determining whether the 2.4GHz remote control link and the 5.8GHz image transmission link appear at the same time period, whether they have synchronous behavior, etc., to further explore cross-frequency collaborative features. The Transformer encoder usually includes multi-head attention layers, feedforward neural networks, layer normalization, etc.

[0099] The processed fusion features are input into the signal classification and discrimination head. The discrimination head is composed of several fully connected layers, and its output is a multi-classification label result, representing the signal type in the current analysis window. The labels can include but are not limited to: legal communication signals (such as WiFi, Bluetooth), unmanned aerial vehicle remote control signals, unmanned aerial vehicle image transmission signals, unknown frequency hopping signals, industrial control signals, etc. The classification result also outputs the corresponding confidence value, which is used for subsequent behavior modeling and credibility evaluation steps.

[0100] In the training phase, the frequency spectrum fusion model can use labeled data sets for supervised learning. The training samples can come from public spectrum signal databases, real captured unmanned aerial vehicle remote control scenarios, or simulator generated data. The classification loss function uses a cross-entropy loss function, and the recognition accuracy can reach more than 90%.

[0101] Through the above structural design, the model has the ability to independently analyze and cross-frequency fusion modeling of signals of different frequency bands, effectively improving the recognition accuracy, anti-interference ability and adaptability of the unmanned aerial vehicle detection system to illegal remote control signals in complex environments. Especially in the context of the increasing number of new communication strategies such as multi-band synchronous frequency hopping and camouflage communication, the model provides an extensible and migratable intelligent spectrum analysis capability, which is a core component module for realizing an efficient and intelligent unmanned aerial vehicle identification process.

[0102] In the embodiment, the behavior modeling module comprises:

[0103] a set of state sequence generation units for converting signal trajectories into state vector sequences;

[0104] an LSTM-based behavior classification network for predicting the behavior type corresponding to the target according to the state sequence;

[0105] The behavior type includes sustained hovering, low-speed cruising, variable-speed return, or short-time interference burst mode.

[0106] The behavior modeling module comprises three functional units, namely, a state sequence generation unit, an LSTM (Long Short-Term Memory Network) behavior classification network, and a behavior type definition set, which work together to form the core process of target signal behavior discrimination.

[0107] First, the state sequence generation unit receives the frequency spectrum trajectory sequence constructed in step S4, extracts and organizes the following state feature fields as part of the state vector, including but not limited to:

[0108] frequency center value (carrier frequency); frequency hopping amplitude; frequency bandwidth (channel width); signal power level; symbol rate and modulation mode; duration and timestamp of each time window.

[0109] After normalization and time series encoding, these fields form a multi-dimensional state sequence. For example, a certain unmanned aerial vehicle jumps over 3 frequency bands in 10 seconds, its symbol rate increases from 500kbps to 1Mbps, and it shows a trend of continuously increasing power, then this behavior state can be abstracted as a state vector time sequence .

[0110] Next, the state sequence is input into a trained LSTM neural network model for behavior classification. The LSTM network is a recurrent neural network with memory capability, which is particularly suitable for processing implicit pattern changes in continuous time series signals. In this invention, the LSTM network can distinguish whether there is a "flight control type behavior" by learning the dynamic characteristics of the frequency trajectory, the modulation mode switching rhythm, and the power evolution law in the target signal state sequence.

[0111] The output of the LSTM network is a multi-classification label result and a corresponding confidence score, and the output label is consistent with the set of behavior type definitions. The set of behavior types can include, but is not limited to:

[0112] Take-off phase (frequency first appears and power rises rapidly); hovering phase (frequency band is stable, modulation is unchanged, and power is stable); cruising phase (modulation mode remains unchanged, and frequency or power slightly fluctuates); return phase (frequency jumps in reverse order, and power gradually decreases); interference frequency hopping phase (short-time continuous switching of frequency bands, and rapid change of power); non-structured behavior (noise signal, environmental interference, etc.).

[0113] In the training process, the signal trajectory sample corresponding to the existing unmanned aerial vehicle control behavior can be used as a training data set to supervise the training of the LSTM model, so that the LSTM model has the ability to identify the typical time sequence signal characteristics in various control chains.

[0114] In addition, the model can also output the confidence based on the confidence, add a reliability label to the behavior recognition result, and pass it to the risk assessment module to calculate whether the target is an abnormal behavior carrier, and improve the accuracy of the system's linkage response.

[0115] In this embodiment, not only can the type of a signal be identified, but the target "what is being done" can also be further judged, jumping from "static identification" to "dynamic understanding" level, providing the system with the ability basis for task-oriented mode identification and intelligent disposal. Especially when facing new protocols, non-standard communication equipment or frequency hopping encryption behavior, the behavior modeling mechanism can discover risks in advance through behavior pattern abnormality judgment, making up for the short board of traditional frequency point and modulation identification technology.

[0116] In this embodiment, the TDOA positioning module comprises:

[0117] A distributed node time synchronization mechanism supports the use of multiple synchronization protocols, including but not limited to GPS PPS pulse, IEEE 1588 PTP protocol, and local crystal oscillator taming mechanism;

[0118] Each receiving node has an independent time verification module that can dynamically calibrate the sampling time deviation;

[0119] The nodes periodically broadcast synchronization state, spectrum sensing state and network health information to judge the current synchronization quality;

[0120] The TDOA positioning algorithm module receives the same signal timestamps, spatial positions and signal strengths uploaded by at least three nodes, and estimates the signal source coordinates using a weighted positioning algorithm;

[0121] The weighting algorithm adjusts the contribution weight of each node according to the synchronization error estimate, the signal-to-noise ratio level, and the distance;

[0122] The positioning result is provided with a confidence level field, and a track point can be output for path backtracking or track splicing analysis.

[0123] In the embodiment, the method further comprises:

[0124] S8, a spectrum behavior fingerprint database is constructed, which is used to record the carrier variation characteristics, modulation mode, frequency hopping mode, symbol rate interval attributes of historical unmanned aerial vehicle communication signals;

[0125] S9, match the spectrum characteristics of the current identified target signal with the fingerprint database, if the matching is successful, mark it as a known manufacturer model or protocol family, if not matched, mark it as a new target.

[0126] The spectrum behavior fingerprint database is composed of a plurality of fingerprint template entries, each entry corresponds to a known communication protocol, unmanned aerial vehicle model or manufacturer equipment. Each fingerprint template includes but is not limited to the following fields:

[0127] Carrier frequency usage mode: including communication channel frequency range, frequency hopping mode, center frequency drift curve;

[0128] Modulation mode characteristics: record the modulation type (such as QPSK, GFSK, OFDM, etc.) adopted by the target, modulation switching characteristics;

[0129] Symbol rate and variation range: including the main communication symbol rate and fluctuation interval (such as 500kbps ± 5%);

[0130] Frequency hopping behavior: including hopping interval, sequence entropy, switching mode (pseudo-random, polling);

[0131] Signal power spectrum density curve: used to distinguish high-power image transmission from low-power control link;

[0132] Time domain activation period and duty cycle: reflect whether the target communication protocol has a control pattern in periodicity;

[0133] Manufacturer identification or communication protocol family label: such as DJI OcuSync, SIYI, WIFI-OFDM, ad hoc 915M protocol, etc.

[0134] During system operation, whenever step S4 identifies a suspected target signal, its spectrum characteristic data (including frequency offset trajectory, modulation mode, carrier frequency, etc.) will be extracted into a target behavior fingerprint vector. The vector is compared with all fingerprint templates in the database, which includes the following steps:

[0135] Normalization: Normalize each dimension of the target fingerprint vector to eliminate device sampling bias.

[0136] Distance calculation: Calculate the similarity score between the target fingerprint and each template using algorithms such as cosine similarity, Euclidean distance, or DTW (Dynamic Time Warping).

[0137] Multi-factor weighted matching: Assign weights to each matching item based on the importance of the fields (e.g., frequency hopping behavior has higher priority than power density), and then calculate the comprehensive score.

[0138] Matching result determination: If the score is higher than the set matching threshold, it is determined that the target is the known model or protocol family represented by the matching template, otherwise it is marked as an unknown target.

[0139] After successful matching, the target signal is labeled with an "ownership tag" such as "DJI Mini 3 image transmission link", and is updated synchronously to the central database for command platform identification. If the matching fails, but the target behavior has obvious pattern structure (such as frequency hopping + timing transmission), the system can automatically generate a new fingerprint entry for it in the database and enter the "gray list" for future further manual annotation.

[0140] In addition, to cope with version updates of communication manufacturers or drone modification behaviors, the fingerprint database supports a dynamic maintenance mechanism. Old fingerprint entries can be set to "frozen" state when they have not been matched for a long time, and their template vectors can be updated by background training if necessary.

[0141] This spectrum behavior fingerprint library mechanism can achieve: rapid label identification of known drones, assisting in improving the generalization ability of the identification model; providing label reference for new target signals, improving the efficiency of manual review; constructing an evolution fingerprint chain for long-term unauthorized communication behavior, assisting in the construction of "black flight database"; as an input source, for cross-modal analysis and calling by behavior modeling module and risk assessment module.

[0142] Therefore, this mechanism significantly enhances the ability of the scheme to close the loop between "spectrum recognition" and "entity attribution recognition", providing key data layer support for building city-level and regional low-altitude airspace intelligent supervision platforms.

[0143] In this embodiment, the calculation of the risk level includes:

[0144] The similarity of the current target signal to the historical legal list;

[0145] The duration of the signal and the frequency of activity;

[0146] Whether the space area covered by the signal trajectory contains a no-fly zone or a warning line;

[0147] Whether a high-risk behavior pattern is identified, specifically including: multi-band synchronous diagram transmission or band hopping encryption behavior.

[0148] In the embodiment, the behavior credibility score field for evaluating the reliability of target signal identification is included in the detection event record, and the score is generated by the following sub-modules:

[0149] The model discriminant confidence module outputs the neural network softmax or sigmoid result of the identification label as the basic score;

[0150] The signal trajectory continuity evaluation module calculates a continuity score based on factors such as frequency trajectory smoothness, modulation mode stability, and duration length;

[0151] The spatial consistency judgment module performs trend consistency matching on the target cross-node positioning path;

[0152] The behavior model fitting degree module evaluates whether the target time series features can be highly fitted and predicted by existing behavior models;

[0153] The score results of each dimension are calculated according to a preset weight to obtain a total credibility score, and a confidence interval is attached, which is used as a reference for event handling decisions.

[0154] In the embodiment, the multiple spectrum receiving nodes construct a distributed spectrum collaborative sensing system, including:

[0155] Each receiving node has a local spectrum preliminary screening module for labeling the identified target signal with a confidence label and a positioning timestamp;

[0156] The identification result is uploaded to the center node through a message queue or a lightweight publish-subscribe protocol, supporting breakpoint continuation;

[0157] The center node constructs an identification graph structure with nodes as vertices and frequency similarity and time sequence proximity as edges;

[0158] Based on the graph neural network, the identification results in the graph are fused and optimized to generate a unified detection label and a consistency score;

[0159] The center node can reversely schedule the front-end node to resample or adjust the frequency band monitoring parameters, realizing adaptive resource allocation.

[0160] In the embodiment, the method further includes:

[0161] S10, the trajectories of the targets identified by different receiving nodes in multiple geographic areas are spliced and analyzed, and if the signal trajectories have band consistency, frequency hopping feature similarity, and spatial continuity, it is determined that they are cross-regional detection results of the same unmanned aerial vehicle target, and are recorded.

[0162] In this embodiment, for example: the national or city-level deployment is divided into multiple monitoring sub-regions, and a set of spectrum receiving nodes and local analysis processing centers are deployed in each region. These regional processing centers have independent running capability and can form target identification events independently according to the whole detection process described in the scheme based on local received data.

[0163] To realize unified identification of cross-regional targets, a cross-regional trajectory splicing module is deployed in the main scheduling platform. The module periodically receives detection events reported from various regions, including but not limited to:

[0164] Target label and identification timestamp; signal frequency band, modulation mode, symbol rate, and other spectrum characteristics; behavior pattern judgment result; positioning information and trajectory segment coordinate point sequence; risk level and credibility score.

[0165] When multiple regional target events show strong correlation in time, frequency band, spatial path, and other dimensions, the system determines them as multiple records of the same target across regions, and performs a trajectory splicing process.

[0166] The process includes the following steps: constructing a time series for all detection events and setting a candidate splicing window (such as ±10 minutes before and after); judging the frequency band consistency (for example, both events are in the 915 MHz frequency band and the modulation mode is QPSK); evaluating the trajectory spatial extension direction and speed vector to determine whether the two targets have “succession”; if the preliminary conditions are met, the two events are listed as splicing candidates.

[0167] Compare whether each segment of frequency offset trajectory, power distribution, and symbol rate stability is highly close; use DTW (dynamic time warping), Frechet distance, and trajectory shape similarity algorithms for time axis alignment; introduce behavior model matching degree as a third-party constraint term to judge whether the two trajectories have the same behavior structure (such as “hovering after returning” splicing “circling transfer”).

[0168] According to the above similarity indexes, score the splicing pairs, and set a minimum splicing confidence threshold (such as 85%); target trajectory pairs exceeding the threshold are marked as “splicable” and merged into a unified target identification (Global Target ID); the spliced trajectory data adopts a time axis fusion strategy for smooth transition processing and reconstruction of continuous spatial paths.

[0169] The merged target is assigned a globally unique identification and recorded as a “long trajectory target”; when displayed on the platform, different colors or segment markers are used to display its behavior of entering, staying, and leaving each region; it can be connected with the air traffic control system to form a “black flight trajectory playback” for later evidence collection and violation identification.

[0170] The mechanism is suitable for scenarios such as cross-region flight of unmanned aerial vehicles, border patrol, city-level air traffic control joint defense and control, etc.; has engineering scalability of "distributed recognition + centralized trajectory integration"; supports offline data reconstruction, and is also suitable for splicing and backtracking of time-delay uploaded data; can be further used for large model training to build behavior templates and analyze brand flight control strategy differences, etc.

[0171] The cross-region trajectory splicing mechanism described in this embodiment establishes a unified target recognition logic among multiple regions and multiple nodes, greatly improving the overall perception consistency and data structure closed loop capability of the system under large-scale deployment, and is one of the core keys to support the evolution of the present scheme to a city-level and national-level unmanned aerial vehicle detection platform.

[0172] In one embodiment, a method for detecting unmanned aerial vehicles based on multi-band radio spectrum analysis is proposed. A large city stadium is deployed and tested as the core area, combined with multi-node synchronous reception, deep learning recognition model, behavior reasoning, risk judgment and cross-region tracking, etc. A complete distributed intelligent detection system is constructed.

[0173] In actual deployment, 8 spectrum receiving nodes are installed on the roofs around the stadium, and the nodes are kept in nanosecond-level time synchronization through the Beidou PPS clock signal. The radio frequency front end of each node has 5-channel parallel sampling capability, covering 2.4GHz, 5.8GHz, 433MHz, 915MHz and 1.2GHz frequency bands respectively. The channels are processed in parallel by the FPGA module, and the band-pass filtering, automatic gain adjustment, IQ sampling and channel marking are completed locally.

[0174] The received original spectrum data generates a short-time Fourier transform (STFT) time-frequency spectrum every 100ms, and the spectrum of multiple frequency bands is input into the fusion analysis model deployed in the central server in the form of channel input. The model includes 5 CNN convolution branches, a multi-head self-attention Transformer module and a signal classification output layer. In the training stage, the model has been supervised by spectrum images containing real remote controllers, legal WiFi, industrial Bluetooth and image transmission interference signals.

[0175] At 16:12 on a certain afternoon, the system identified a group of signals with frequency hopping characteristics on the 915MHz frequency band, and the model output label was "suspected image transmission link" with a confidence of 92%. The signal was received by multiple nodes at the same time, and the TDOA positioning module was used to reverse the source location to about 30 meters above the southwest stand of the stadium.

[0176] Subsequently, the system sends the continuous track of the signal to the behavior modeling module for pattern recognition. The LSTM behavior classification network identifies its behavior chain as "short hovering - moving at a constant speed north - turning northwest - hovering - rapid falling", inferring the "take-off - directional shooting - return" typical behavior pattern, with a model fitting degree of 87%. Combined with the characteristics of high signal duty cycle, strong frequency hopping law, stable communication power, etc., the target is judged to have clear flight control intention.

[0177] At the same time, the system calculates the credibility score of the target signal, with a model confidence of 92%, a track continuity score of 88%, a spatial consistency score of 90%, and a behavior fitting score of 87%. The final comprehensive credibility score is 89.2, with a confidence interval of ±3.1, belonging to a high credibility target.

[0178] The system automatically starts risk level calculation, and according to the characteristics that the target is located in the temporary no-fly zone of the event, uses illegal frequency hopping channel, and has obvious image transmission behavior, the final risk score is 92, and it is marked as "high-risk target". The platform immediately issues a warning and links to the nearby wireless suppression equipment to start the preset interference strategy.

[0179] More importantly, within the next 30 minutes, the target was again identified by another subsystem over a certain community 2.6 kilometers northwest of the main venue. The identification system belongs to the same city's distributed defense network. The system compares the two target data through frequency band, behavior pattern, carrier frequency stability, image transmission rate, etc. fields, and through the DTW alignment in the track splicing module and the spatial curvature consistency analysis, it is confirmed that it is the same target. Finally, the system merges the target into "cross-region target UAV-20250807114", reconstructs its complete track and archives the evidence.

[0180] As can be seen from the above embodiment, the method has good recognition accuracy, positioning accuracy and target tracking ability in a multi-band complex communication environment. Especially in the face of frequency hopping signals, non-standard image transmission, cross-region black flight and other scenarios, the system realizes rapid, reliable and efficient aircraft detection and discrimination through multi-modal recognition and collaborative perception ability, effectively supporting the active supervision needs of city-level low-altitude airspace.

[0181] The technical features of the above embodiments can be combined in any way. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the present disclosure.

Claims

1. A method for detecting unmanned aerial vehicles based on multi-band radio spectrum analysis, characterized in that, The method comprises: S1, laying a plurality of spectrum receiving nodes in a region to be monitored, wherein the plurality of nodes have multi-band radio collection capability for collecting a plurality of frequency band signal data including 2.4 GHz, 5.8 GHz, 433 MHz, 915 MHz and 1.2 GHz in parallel; S2, performing noise reduction, normalization and frame window segmentation processing on the collected signals of each frequency band to generate a multi-channel time-frequency spectrum graph in a unified format; S3, inputting the multi-channel time-frequency spectrum graph into a spectrum fusion analysis model containing convolution structure and Transformer structure, extracting local features within the frequency band and time sequence features between the frequency bands respectively, and outputting target discrimination results containing confidence and labels; S4, extracting target carrier frequency, symbol rate, frequency offset trajectory and modulation mode signal features for the target discriminated as suspected unmanned aerial vehicle control signal or image transmission signal, and constructing a time series signal trajectory; S5, sending the target trajectory to a behavior modeling module to model the target spectrum behavior through a state transition analysis method for identifying suspected flight control behavior, the control behavior including hovering, cruising and homing modes; S6, calculating two-dimensional or three-dimensional positioning information of the signal source based on the timestamps of the same target signal received by the plurality of spectrum receiving nodes, and realizing target motion path reconstruction in combination with historical trajectory data; S7, calculating the risk level of the identified target signal, and generating an unmanned aerial vehicle detection event record containing target spatial position, signal category, trajectory features and threat level.

2. The method of claim 1, wherein, The radio frequency front end of the spectrum receiving node comprises: A multi-channel parallel radio frequency receiving link, each link being configured with a low noise amplifier, a variable bandwidth bandpass filter, an automatic gain control unit, a programmable local oscillator frequency source and an analog-to-digital conversion module; The bandpass filter supports on-demand switching of center frequency and bandwidth parameters for dynamically adapting different frequency band signals; The local oscillator frequency source and the analog-to-digital conversion module are both provided with a unified reference clock by a high-precision temperature-compensated crystal oscillator; Data sampling synchronization and timestamp injection are realized between the multiple links through an FPGA logic controller, and docking with GPS, Beidou PPS pulse or IEEE 1588 time protocol is supported; The collected spectrum data is packaged in IQ pair format with channel number, frequency band identification and high-precision timestamp fields for subsequent central analysis processing.

3. The method of claim 1, wherein, The spectrum fusion analysis model comprises: At least one parallel convolution branch for different frequency band time-frequency spectrum graphs for extracting local texture features and edge features of the spectrum graph; A Transformer encoder module with a multi-head self-attention structure for capturing time evolution rules and cross-frequency band correlation; A classification discrimination head for outputting multi-class label results including legal communication, unmanned aerial vehicle remote control, image transmission channel and unknown signal.

4. The method of claim 1, wherein, The behavior modeling module comprises: A group of state sequence generation units for converting signal trajectories into state vector sequences; An LSTM-based behavior classification network for predicting the behavior type corresponding to the target according to the state sequence; The behavior types include sustained hovering, low-speed cruising, variable-speed return, or short-time interference burst mode.

5. The method of claim 1, wherein, The TDOA positioning module comprises: A distributed node time synchronization mechanism supports the use of multiple synchronization protocols, including but not limited to GPS PPS pulse, IEEE1588 PTP protocol, and local crystal oscillator taming mechanism; Each receiving node has an independent time calibration module that can dynamically calibrate the sampling time deviation; Periodically broadcast synchronization status, spectrum sensing status, and network health information between nodes to determine the current synchronization quality; The TDOA positioning algorithm module receives the same signal timestamps, spatial positions, and signal strengths uploaded by at least three nodes, and estimates the signal source coordinates using a weighted positioning algorithm to obtain the positioning result; The weighted algorithm adjusts the contribution weight according to the synchronization error estimate of each node, signal-to-noise ratio level, and distance; the positioning result has a confidence level field and can output trajectory points for path backtracking or trajectory stitching analysis.

6. The method of claim 1, wherein, The method further comprises: S8, constructing a spectrum behavior fingerprint database to record the carrier variation characteristics, modulation method, frequency hopping mode, and symbol rate interval attributes of historical unmanned aerial vehicle communication signals; S9, matching the spectrum characteristics of the current identified target signal with the fingerprint database, if the matching is successful, marking it as a known manufacturer model or protocol family, if not, marking it as a new target.

7. The method of claim 1, wherein, The calculation basis of the risk level includes: The similarity of the current target signal and the historical legal list; The duration and frequency activity level of the signal; Whether the space area covered by the signal trajectory contains a no-fly zone or warning line; Whether a high-risk behavior mode is identified, specifically including: multi-band synchronous image transmission or frequency band hopping encryption behavior.

8. The method of claim 1, wherein, The behavior confidence score field in the detection event record is used to evaluate the reliability of target signal identification, and the score is generated by the following sub-modules: The model discrimination confidence module outputs the neural network softmax or sigmoid result of the identification label as the basic score; The signal trajectory continuity evaluation module calculates the continuity score based on factors such as frequency trajectory smoothness, modulation stability, and duration length; The spatial consistency judgment module matches the trend consistency of the target cross-node positioning path; The behavior model fitting degree module evaluates whether the target timing characteristics can be highly fitted and predicted by existing behavior models; The total confidence score is calculated according to the preset weight of each dimension, and the confidence interval is attached, which is used as a reference for event handling decisions.

9. The method of claim 1, wherein, The multiple spectrum receiving nodes construct a distributed spectrum collaborative sensing system, comprising: Each receiving node has a local spectrum preliminary screening module that labels the identified target signal with a confidence label and a positioning timestamp; The identification result is uploaded to the center node through the message queue or lightweight publish-subscribe protocol, supporting breakpoint resume; The center node constructs an identification graph structure with nodes as vertices and frequency similarity and timing proximity as edges; Based on the graph neural network, the identification results in the graph are fused and optimized to generate a unified detection label and consistency score. The central node can reverse schedule the front-end node to resample or adjust the frequency band monitoring parameters, to realize adaptive resource allocation.

10. The method of claim 1, wherein, The method further comprises: S10, the trajectory splicing analysis is carried out to the target respectively identified by different receiving nodes in multiple geographical areas, if the signal trajectory has frequency band consistency, frequency hopping characteristic similarity and spatial continuity, it is determined that it is the cross-region detection result of the same unmanned aerial vehicle target, and is combined and recorded.

Images