UAV detection method and system based on radio intelligent fusion

Abstract

This application relates to the field of data processing technology and discloses a drone detection method and system based on radio intelligent fusion. The method includes: acquiring radio signals to obtain I / Q data; performing time-frequency conversion and feature extraction; inputting into a convolutional neural network for identification; applying a clustering joint recognition algorithm to determine the type; constructing a relay network for fusion processing to determine location and identity; and calculating a threat index to implement early warning classification and form a defense strategy. This application can achieve collaborative perception and intelligent fusion of multi-source information in complex electromagnetic environments through a multi-level relay network, improving the accuracy, coverage, and real-time performance of drone detection, and implementing adaptive early warning and defense responses based on intelligent threat assessment.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method and system for detecting unmanned aerial vehicles (UAVs) based on radio intelligent fusion. Background Art

[0002] In the field of drone detection and defense, existing technologies primarily rely on single detection methods, such as radar detection, radio monitoring, or optoelectronic tracking. Radar detection technology utilizes the principle of electromagnetic wave reflection and can detect drone targets at relatively long distances, but it faces challenges such as difficulty identifying small targets and susceptibility to interference. Radio monitoring technology can detect the presence of drones within a certain range by receiving communication signals between the drone and the remote controller, but its effectiveness against drones with encrypted communications or autonomous flight is limited. Optoelectronic tracking technology relies on infrared or visible light cameras to capture drone images, offering the advantage of intuitive visualization, but is significantly limited by environmental conditions such as weather and lighting. Furthermore, existing drone identification methods are typically based on simple feature matching or threshold judgment, resulting in low recognition accuracy and a lack of effective multi-source information fusion mechanisms, making it difficult to address the need for precise identification in complex electromagnetic environments.

[0003] The main shortcomings of existing technologies are: First, the limited perception capabilities of a single sensor prevent it from fully capturing the characteristic information of drones, resulting in high false alarm and missed alarm rates. Second, traditional signal processing methods are insufficient for processing drone signals with weak signals and low signal-to-noise ratios, making it difficult to reliably detect drones in complex environments. Third, there is a lack of effective multi-sensor collaborative detection mechanisms, with each sensor operating independently and failing to fully utilize information complementarity. Finally, existing drone threat assessments are mostly static models, lacking intelligent dynamic assessment and early warning response mechanisms, and are slow to respond to rapidly changing drone threats. These shortcomings severely restrict the effectiveness of drone detection systems in important application scenarios such as no-fly zone monitoring and critical infrastructure protection. Summary of the Invention

[0004] This application provides a drone detection method and system based on radio intelligent fusion, which can realize collaborative perception and intelligent fusion of multi-source information through a multi-level relay network in a complex electromagnetic environment, improve the accuracy, coverage and real-time performance of drone detection, and realize adaptive early warning and defense response based on threat intelligent assessment.

[0005] In the first aspect, the present application provides a drone detection method based on radio intelligent fusion, which includes: collecting radio signals through distributed relay sensor nodes to obtain digital I / Q data; performing time-frequency domain conversion and feature extraction on the digital I / Q data to obtain a drone radio signal feature vector set; inputting the drone radio signal feature vector set into a convolutional neural network for training and recognition to obtain a drone signal recognition result; obtaining a drone type recognition result by clustering the drone signal recognition result through a signal joint recognition algorithm; constructing a multi-level relay network based on the drone type recognition result to perform collaborative data fusion processing to obtain the drone location and identity information; calculating the threat index for the drone location and identity information and performing early warning classification based on the relay network to obtain a defense response strategy.

[0006] In a second aspect, the present application provides a drone detection system based on radio intelligent fusion, the drone detection system based on radio intelligent fusion comprising:

[0007] The acquisition module is used to collect radio signals through distributed relay sensor nodes to obtain digital I / Q data;

[0008] An extraction module, configured to perform time-frequency domain conversion and feature extraction on the digitized I / Q data to obtain a set of UAV radio signal feature vectors;

[0009] An input module, configured to input the drone radio signal feature vector set into a convolutional neural network for training and recognition, thereby obtaining a drone signal recognition result;

[0010] an identification module, configured to identify the drone signal using a clustering signal joint identification algorithm to obtain a drone type identification result;

[0011] A fusion module is used to construct a multi-level relay network based on the UAV type identification results to perform collaborative data fusion processing to obtain the UAV location and identity information;

[0012] The classification module is used to calculate the threat index of the drone's location and identity information and perform early warning classification based on the relay network to obtain a defense response strategy.

[0013] In a third aspect, a computer device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned drone detection method based on radio intelligent fusion.

[0014] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, enables the computer to execute the above-mentioned drone detection method based on radio intelligent fusion.

[0015] In the technical solution provided by this application, radio signals are collected by distributed relay sensor nodes to obtain digital I / Q data, thereby achieving seamless coverage of a wide area and solving the problem of limited coverage of a single sensor. The digital I / Q data is converted into time and frequency domains and features are extracted to obtain a set of feature vectors of drone radio signals, which fully exploits the multi-dimensional features of drone signals and improves the richness and discrimination of feature expression. The set of feature vectors of drone radio signals is input into a convolutional neural network for training and recognition to obtain drone signal recognition results, so that the model can automatically learn the deep feature patterns of the signal, and the recognition ability is no longer limited to manually preset feature rules. The preliminary signal is jointly identified by a clustering signal joint recognition algorithm. The drone signal recognition results are processed to obtain the drone type recognition results. This clever integration of cluster analysis and pattern recognition technology can accurately distinguish drone models with similar characteristics, significantly reducing the misidentification rate. Based on the drone type recognition results, a multi-level relay network is constructed to perform collaborative data fusion processing to obtain the drone location and identity information, solving the information island problem in traditional single-node detection and achieving effective complementarity and redundant verification of information between different nodes. The threat index is calculated for the drone location and identity information and an early warning classification based on the relay network is performed to obtain a defense response strategy, building a complete intelligent defense chain from perception to decision-making, ensuring that corresponding defense measures are taken against drones of different threat levels. The present invention applies artificial intelligence algorithms and models in specific drone detection application fields, especially convolutional neural networks and clustering signal joint recognition algorithms. It not only surpasses traditional methods in the feature extraction and recognition stages, but also significantly improves detection accuracy and anti-interference ability through algorithm features, and shows excellent adaptability in complex electromagnetic environments; the collaborative data fusion mechanism of the multi-level relay network fully considers the spatial characteristics of radio signal propagation, and utilizes the spatial diversity formed by distributed sensing nodes to maintain stable and reliable detection effects in strong interference environments; the adaptive early warning classification strategy based on threat index deeply combines artificial intelligence decision-making technology with professional field knowledge to form an intelligent defense system against drone threats. Compared with traditional static defense strategies, it has higher targeting and resource utilization efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 This is a schematic diagram of an embodiment of a UAV detection method based on radio intelligent fusion in an embodiment of the present application;

[0018] Figure 2 This is a schematic diagram of an embodiment of a UAV detection system based on radio intelligent fusion in an embodiment of the present application;

[0019] Figure 3 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The embodiments of the present application provide a method and system for detecting unmanned aerial vehicles based on radio intelligence fusion. The terms "first," "second," "third," "fourth," and so on (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way are interchangeable where appropriate so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0021] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of the drone detection method based on radio intelligent fusion includes:

[0022] Step S101: Collect radio signals through distributed relay sensor nodes to obtain digital I / Q data;

[0023] Step S102: performing time-frequency domain conversion and feature extraction on the digitized I / Q data to obtain a set of UAV radio signal feature vectors;

[0024] Step S103: Input the UAV radio signal feature vector set into the convolutional neural network for training and recognition to obtain the UAV signal recognition result;

[0025] Step S104: using a clustering signal joint recognition algorithm to identify the drone signal and obtain a drone type recognition result;

[0026] Step S105: construct a multi-level relay network based on the drone type identification result to perform collaborative data fusion processing to obtain the drone location and identity information;

[0027] Step S106: Calculate the threat index based on the drone's location and identity information and perform early warning classification based on the relay network to obtain a defense response strategy.

[0028] It is understandable that the execution subject of this application can be a UAV detection system based on radio intelligent fusion, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0029] Specifically, radio signals are collected through distributed relay sensor nodes to obtain digitized I / Q data. Distributed relay sensor nodes are radio signal receiving devices deployed at different geographical locations. These devices form a network that can capture radio signals in the environment. The devices contain general-purpose software radio peripherals and corresponding signal processing modules, operating in the 2.4GHz ISM band, 5.8GHz band, and 900MHz band, which are the primary operating frequency bands for the remote control, image transmission, and telemetry systems of most commercial drones. During the signal acquisition process, the sensor nodes pre-amplify the received analog RF signal using a low-noise amplifier. They then use a bandpass filter to filter out out-of-band interference signals. The RF signal is then mixed with a carrier signal generated by a local oscillator through a mixer and down-converted to an intermediate frequency or baseband signal. The signal is then digitized into a complex I / Q data stream using an analog-to-digital converter at a sampling rate of at least 20MHz.

[0030] The digitized I / Q data is converted to the time-frequency domain and features are extracted to obtain a set of drone radio signal feature vectors. The I / Q data is converted to the time-frequency domain using a short-time Fourier transform (SFT) to generate a signal spectrogram. Parameters such as power spectral density and frequency occupied bandwidth are calculated from the signal spectrogram to obtain frequency domain feature parameters. Envelope analysis is also performed on the I / Q data to extract time domain features, including signal envelope shape and pulse repetition interval. High-order cumulants and cyclic spectrum features are then calculated from the time and frequency domain feature parameters to obtain modulation identification parameters, which are used to identify the modulation type used by the drone signal. MAC layer feature extraction is performed on the modulation identification parameters and signal message structure to obtain communication protocol features, including frame structure and preamble pattern. Finally, all extracted features are normalized and weighted to form a set of drone radio signal feature vectors.

[0031] The drone radio signal feature vector set was input into a convolutional neural network for training and recognition, resulting in drone signal recognition results. The convolutional neural network architecture consists of five convolutional layers, each followed by a max pooling layer and a batch normalization layer. The first convolutional layer uses 32 3×3 convolution kernels, the second layer uses 64 3×3 convolution kernels, the third layer uses 128 3×3 convolution kernels, the fourth layer uses 256 3×3 convolution kernels, and the fifth layer uses 512 3×3 convolution kernels. Two fully connected layers were designed after the convolutional layers, containing 1024 and 512 neurons, respectively. Training was performed using the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 200 epochs. The feature vector set of drone radio signals is first normalized to conform to the input format of a convolutional neural network. It is then fed into the convolutional layer for feature mapping, generating a hierarchical feature representation. After dimensionality reduction and normalization in the pooling and batch normalization layers, the signal is fed into the fully connected layer for nonlinear transformation. Finally, the probability distribution of each drone type is calculated using the Softmax function, and the category with the highest probability is selected as the recognition result. The drone signal recognition results are processed using the clustered signal joint recognition algorithm to obtain the drone type recognition result. The clustered signal joint recognition algorithm first configures a two-dimensional self-organizing map network with a network size of 10×10 and 100 neurons. Training is performed in batch mode, with an initial learning rate of 0.5 and exponential decay, and an initial neighborhood radius of 5 and gradually decreasing. Known drone samples are fed into the trained network, and the distribution of each type of sample on the grid is statistically analyzed to form a grid location probability distribution map. The drone signal recognition result is then fed into the network, and the distance to each neuron is calculated. The neuron with the smallest distance is identified as the matching unit, resulting in the signal's location in feature space. This location is then queried from the grid location probability distribution map to calculate the spatial mapping probability value. The spatial mapping probability value and the classification probability output by the convolutional neural network are weighted and fused according to the dynamically adjusted weight coefficient to obtain the comprehensive recognition probability. The recognition results are then smoothed in time series using the state transition matrix to obtain the drone type recognition result.

[0032] Based on the drone type identification results, a multi-level relay network is constructed to perform collaborative data fusion processing to obtain the drone's location and identity information. The multi-level relay network divides sensor nodes into master nodes, relay nodes, and edge nodes, establishing inter-node communication links through a hierarchical network topology. Each node is synchronized using a precision time protocol, with clock deviation controlled to within 100 nanoseconds. A trilateration algorithm is used to calculate the signal source location coordinates based on each node's geographic location and signal reception strength. The feature vector extracted from each node is transmitted to the upper-level node through the network, where it is weighted averaged and dimensionality reduced to obtain a fused feature vector. The confidence and uncertainty of each node's independent identification results are calculated using an evidence theory framework, and the weight coefficients are dynamically adjusted to obtain a comprehensive identification result. The signal source location coordinates are correlated and matched with the comprehensive identification results, and trajectory correlation is performed using a tracking filter to obtain the drone's location and identity information.

[0033] A threat index is calculated based on the drone's location and identity information, and a relay network-based early warning classification is applied to determine a defense response strategy. The threat index calculation comprehensively considers factors such as the risk factor of the drone type, its proximity to sensitive areas, its size, its approach speed, and the safety level of the sensitive areas. The threat index is categorized into four levels: normal (0-0.25), caution (0.25-0.5), warning (0.5-0.75), and emergency (0.75-1.0). The frequency of scanning the monitored area is adjusted based on the early warning level, and notifications of varying priorities are sent to relevant personnel. Target confirmation is performed by activating the corresponding detection equipment through a multi-level relay network, generating target confirmation information. The drone's trajectory is tracked and predicted, resulting in a trajectory prediction result. Response resources are allocated based on the target confirmation information, trajectory prediction results, and early warning level determination, forming a defense response strategy. In a certain actual test, when a drone approached a sensitive area, the system calculated a threat index of 0.68, which was judged to be a warning level. The photoelectric detection equipment was immediately activated for confirmation and predicted that it would enter the core area in 10 seconds. Therefore, the electromagnetic interference defense equipment was triggered, successfully preventing the drone from approaching further.

[0034] In the embodiment of the present application, radio signals are collected by distributed relay sensor nodes to obtain digital I / Q data, thereby achieving seamless coverage of a wide area and solving the problem of limited coverage of a single sensor; the digital I / Q data is converted into time and frequency domains and feature extracted to obtain a set of feature vectors of UAV radio signals, which fully exploits the multi-dimensional features of UAV signals and improves the richness and discrimination of feature expression; the set of feature vectors of UAV radio signals is input into a convolutional neural network for training and recognition to obtain a UAV signal recognition result, so that the model can automatically learn the deep feature patterns of the signal, and the recognition ability is no longer limited to the manually preset feature rules; the preliminary UAV signal recognition is performed by a clustering signal joint recognition algorithm. The drone signal recognition results are processed to obtain the drone type recognition results. It cleverly combines cluster analysis and pattern recognition technology, and can accurately distinguish drone models with similar characteristics, significantly reducing the misidentification rate. According to the drone type recognition results, a multi-level relay network is constructed to perform collaborative data fusion processing to obtain the drone location and identity information, solving the information island problem in traditional single-node detection and realizing effective complementarity and redundant verification of information between different nodes. The threat index is calculated for the drone location and identity information and an early warning classification based on the relay network is performed to obtain the defense response strategy, building a complete intelligent defense chain from perception to decision-making, ensuring that corresponding defense measures are taken against drones of different threat levels. The present invention applies artificial intelligence algorithms and models in specific drone detection application fields, especially convolutional neural networks and clustering signal joint recognition algorithms. It not only surpasses traditional methods in the feature extraction and recognition stages, but also significantly improves detection accuracy and anti-interference ability through algorithm features, and shows excellent adaptability in complex electromagnetic environments; the collaborative data fusion mechanism of the multi-level relay network fully considers the spatial characteristics of radio signal propagation, and utilizes the spatial diversity formed by distributed sensing nodes to maintain stable and reliable detection effects in strong interference environments; the adaptive early warning classification strategy based on threat index deeply combines artificial intelligence decision-making technology with professional field knowledge to form an intelligent defense system against drone threats. Compared with traditional static defense strategies, it has higher targeting and resource utilization efficiency.

[0035] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0036] The radio signal in the environment is pre-amplified by a low-noise amplifier to obtain an analog radio frequency signal;

[0037] The analog radio frequency signal is filtered out of the band by a bandpass filter to obtain a filtered radio frequency signal;

[0038] The filtered RF signal is down-converted using a mixer and a local oscillator to obtain an intermediate frequency or baseband signal;

[0039] The intermediate frequency or baseband signal is digitized at a high sampling rate through an analog-to-digital converter to obtain the original I / Q data stream;

[0040] Perform digital downsampling and digital filtering on the original I / Q data stream to obtain noise-reduced I / Q data;

[0041] The noise-reduced I / Q data is marked with a timestamp and geographic location information to obtain digitized I / Q data.

[0042] Specifically, the radio signals in the environment are pre-amplified through a low-noise amplifier to obtain an amplified analog RF signal. A low-noise amplifier is an electronic amplifier whose main function is to amplify weak RF signals while introducing as little noise as possible. In drone detection systems, the received radio signals are usually weak in strength, with signal power ranging from -90dBm to -120dBm. The low-noise amplifier amplifies these weak signals to a level suitable for subsequent processing, usually with a gain between 20dB and 30dB, while maintaining a low noise figure, generally controlled below 2dB. The dynamic range of the amplifier must be considered during the amplification process to ensure that the amplifier is not saturated due to an overly strong signal, or overwhelmed by noise due to a too weak signal.

[0043] The amplified analog RF signal is filtered through a bandpass filter to remove out-of-band interference, resulting in a filtered RF signal. A bandpass filter is an electronic filter that only allows signals within a specific frequency range to pass, blocking signals at other frequencies. In drone detection systems, the bandpass filter is designed to pass only frequency bands commonly used by drones, such as the 2.4 GHz ISM band, the 5.8 GHz band, and the 900 MHz band. The filter bandwidth is set based on the bandwidth characteristics of the target drone's communication signal, typically between 20 MHz and 40 MHz. The filtering process utilizes a multi-stage bandpass filter structure, providing at least 60 dB of out-of-band rejection, effectively reducing the impact of interference signals generated by other electronic devices on the detection system. The filtered signal retains the characteristics of the frequency band likely used by drone communications while significantly reducing interference from other frequency bands. The filtered RF signal is then down-converted using a mixer and a local oscillator to produce an intermediate frequency (IF) or baseband signal. A mixer is a circuit that multiplies the input signal with a local oscillator signal to generate sum and difference frequency signals. The local oscillator generates a stable sinusoidal signal, the frequency of which is set according to the target drone's communication frequency. During the downconversion process, the filtered RF signal is mixed with a local oscillator (LO) signal, converting the high-frequency RF signal to a lower intermediate frequency (IF) signal (typically in the tens of MHz range) or directly to baseband (near zero frequency). For example, when detecting a 2.4 GHz drone signal, the LO frequency can be set to 2.385 GHz. This mixing yields a 15 MHz IF signal for subsequent digital processing. During the downconversion process, the LO's phase noise and frequency stability must be controlled to ensure the quality of the converted signal.

[0044] The intermediate frequency (IF) or baseband signal is digitized at a high sampling rate using an analog-to-digital converter (ADC), producing a raw I / Q data stream. An ADC converts a continuous analog signal into a discrete digital signal. In drone detection systems, the ADC must have a sufficient sampling rate and bit depth to accurately capture signal characteristics. A sampling rate of at least 20 MHz is sufficient to cover the bandwidth of most commercial drone communication signals, as determined by the Nyquist sampling theorem. The ADC bit depth is typically 12 or 16 bits to provide sufficient dynamic range and digitization accuracy. The digitization process converts the IF or baseband signal into in-phase (I) and quadrature (Q) components, or I / Q data. This representation preserves the signal's amplitude and phase information. I / Q data is represented in complex form, with I representing the real part and Q representing the imaginary part, describing the instantaneous state of the signal.

[0045] The raw I / Q data stream undergoes digital downsampling and digital filtering to produce de-noised I / Q data. Digital downsampling involves reducing the signal sampling rate to reduce the data volume and lower the computational complexity of subsequent processing. In drone detection systems, the raw I / Q data stream may be sampled at too high a rate. Downsampling reduces this rate to the required rate, typically to 1 / 2 or 1 / 4 of the original rate. Anti-aliasing filtering is required before downsampling to prevent high-frequency components from folding into low-frequency components, which can cause signal distortion. Digital filtering involves applying bandpass, lowpass, or highpass filtering to the signal in the digital domain to further suppress noise and interference. FIR (finite impulse response) filters are commonly used in drone detection systems because their linear phase characteristics help preserve signal morphology. After digital filtering, the noise content of the I / Q data is significantly reduced, and the signal characteristics are more prominent. The de-noised I / Q data is then tagged with a timestamp and geolocation information to produce digitized I / Q data. The timestamp is the precise time information recording the moment of data acquisition, typically with microsecond or nanosecond accuracy, and is obtained using a GPS receiver or the Network Time Protocol (NTP). Geographic location information, including the longitude, latitude, and altitude of the sensor node, is also acquired via a GPS receiver, achieving meter-level accuracy. The addition of timestamps and geolocation information allows for temporal and spatial correlation of data collected by multiple sensor nodes, providing a foundation for subsequent multi-node collaborative processing and target positioning. During the data tagging process, timestamps and geolocation information are appended to the I / Q packet header as metadata, creating a digitized I / Q packet for easier network transmission and subsequent processing.

[0046] For example, a sensor node receives a suspected drone signal with an initial signal strength of -100dBm. After amplification by a 30dB low-noise amplifier, the signal strength reaches -70dBm. After bandpass filtering, interfering signals outside the 2.4GHz band are attenuated by 60dB, improving the signal-to-noise ratio from 5dB to 25dB. The signal is then mixed with a 2.385GHz local oscillator to generate an intermediate frequency (IF) signal with a center frequency of 15MHz. This IF signal is digitized using a 16-bit, 50MHz sampling rate analog-to-digital converter to produce a raw I / Q data stream. This raw data is downsampled by a factor of 4 to a sampling rate of 12.5MHz and digitally filtered using a 64th-order FIR filter, further improving the signal-to-noise ratio to 30dB. Finally, a microsecond-accurate timestamp and meter-level geolocation are added to complete the preparation of the digitized I / Q data.

[0047] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0048] The digital I / Q data is converted into the time-frequency domain through short-time Fourier transform to obtain the signal spectrum;

[0049] Perform power spectrum density calculation and frequency occupied bandwidth analysis on the signal spectrum to obtain frequency domain characteristic parameters;

[0050] The digital I / Q data is subjected to time domain feature extraction through envelope analysis to obtain time domain features;

[0051] Calculate the high-order cumulative amount and cyclic spectrum characteristics of the time domain and frequency domain characteristic parameters to obtain the modulation recognition parameters;

[0052] Perform MAC layer feature extraction on modulation identification parameters and signal message structure to obtain communication protocol features;

[0053] The frequency domain feature parameters, time domain features, modulation identification parameters and communication protocol features are normalized and weighted to obtain the UAV radio signal feature vector set.

[0054] Specifically, the digitized I / Q data is converted to the time-frequency domain using a short-time Fourier transform (SFT). The SFT is a method for time-frequency analysis of signals that is achieved by windowing the signal and then performing a Fourier transform. The digitized I / Q data is divided into multiple overlapping time segments. Each segment is multiplied by a window function (such as a Hamming window or a Heyman window) and then subjected to a discrete Fourier transform. In drone signal processing, the window length is typically 1024 points with a 50% overlap to balance time and frequency resolution. The Fourier transform results of each time segment form a column of spectra. The spectra of multiple time segments are arranged in chronological order to form a two-dimensional time-spectrum graph, with time on the horizontal axis and frequency on the vertical axis. The color depth indicates signal energy. This signal spectrogram provides a visual representation of the time-varying and frequency characteristics of drone signals.

[0055] The signal spectrum is subjected to power spectral density calculation and frequency occupied bandwidth analysis to obtain frequency domain characteristic parameters. Power spectral density calculation estimates the energy of each frequency point in the signal spectrum. Based on the Welch method of the periodogram, the power spectra calculated over multiple time periods are averaged to reduce the impact of noise. The calculation results show the distribution of signal energy over frequency. Frequency occupied bandwidth analysis measures the effective bandwidth of the signal power spectrum. The specific method is to determine the continuous frequency range where the signal energy exceeds the noise threshold, or to determine the frequency range that contains 90% of the total signal energy. In addition to bandwidth, parameters such as center frequency, frequency offset, and spectral peak ratio are also calculated. For example, the typical bandwidth of the remote control signal of a certain brand of drone is 4MHz, while the bandwidth of the image transmission signal reaches 20MHz, with the center frequency in the 2.4GHz band.

[0056] The digitized I / Q data is subjected to time domain feature extraction through envelope analysis to obtain time domain features. Envelope analysis is a method for obtaining instantaneous amplitude changes of a signal. The signal envelope is obtained by calculating the modulus of the I / Q data or using the Hilbert transform. The time domain features extracted from the signal envelope include the signal's start time, duration, rise time, fall time, pulse repetition interval, duty cycle, etc. It also includes statistical features of the signal amplitude, such as mean, variance, kurtosis, skewness, etc. Time domain features are particularly important for identifying different drone communication signal formats, because drone remote control signals from different manufacturers usually have different signal transmission timing patterns. For example, the control signal of a certain model of drone has a frame duration of 20 milliseconds, a pulse repetition interval of 5 milliseconds, and a rise time of approximately 0.2 milliseconds. These features are significantly different from those of other models of drones.

[0057] High-order cumulants and cyclic spectrum features are calculated for the time-domain and frequency-domain feature parameters to obtain modulation identification parameters. High-order cumulants are high-order statistical properties of the signal that are unaffected by Gaussian noise and are effective for identifying the signal's modulation scheme. Second-order cumulants are equivalent to the autocorrelation function, while third- and fourth-order cumulants reflect the signal's nonlinear characteristics. Cyclic spectrum features characterize the signal's periodicity. By calculating the correlation of the signal at different frequency offsets, a two-dimensional cyclic spectrum is obtained, which can distinguish different modulation types. Modulation identification parameters extracted from the high-order cumulants and cyclic spectrum include spectral peak value, spectral line density, and spectral line distribution, which are used to determine the signal's modulation scheme. Drones from different manufacturers often use different modulation schemes. For example, the remote control signal of a certain brand of drone uses FHSS (frequency hopping spread spectrum) technology with 2FSK modulation, while the image transmission signal uses OFDM modulation.

[0058] MAC layer feature extraction is performed on the modulation identification parameters and signal message structure to obtain communication protocol features. MAC layer feature extraction analyzes the communication protocol structure, including frame format, preamble pattern, packet header structure, and data payload length distribution. By synchronizing and demodulating the signal, data frame boundaries are identified, and then the intra-frame structural features are analyzed. Common drone communication protocols such as MAVLink, SBUS, and DSM have specific frame structures and checksums. The extracted MAC layer features include statistical characteristics of frame length, interframe interval, preamble sequence pattern, and header field characteristics. These features are particularly effective in distinguishing drones from different manufacturers, as even manufacturers using the same physical layer technology often use proprietary communication protocols. For example, a drone from one manufacturer uses a customized communication protocol characterized by a fixed 64-byte frame length, a specific 8-byte preamble sequence, and a 20-ms interframe interval—features that are significantly different from drones from other manufacturers. Frequency domain feature parameters, time domain features, modulation identification parameters, and communication protocol features are normalized and weighted to obtain a set of drone radio signal feature vectors. Normalization converts all features to the same numerical range, typically between [0 and 1]. Common methods include min-max normalization and Z-score normalization. Normalized features can be effectively compared and fused. Weighting is performed by assigning different weights to each feature based on its discriminative power. This is typically determined using feature importance assessment methods such as information gain and the Gini index. The resulting feature vector set is a high-dimensional vector composed of the weighted combinations of the normalized features. Each vector corresponds to a signal sample, and each element of the vector corresponds to an eigenvalue. The entire feature extraction process transforms raw I / Q data into structured feature vectors.

[0059] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0060] Normalize the feature vector set of the drone radio signal to obtain training data that conforms to the input format of the convolutional neural network;

[0061] The training data is fed into a network structure containing five convolutional layers for feature mapping to obtain hierarchical feature expression data;

[0062] The hierarchical feature expression data is subjected to dimensionality reduction and standardization processing through the maximum pooling layer and batch normalization layer to obtain compressed feature expression data;

[0063] The compressed feature expression data is input into two fully connected layers for nonlinear transformation to obtain vector representation data in high-dimensional feature space;

[0064] The vector representation data in the high-dimensional feature space is used to calculate the probability distribution through the Softmax function to obtain the probability distribution of each drone type;

[0065] Based on the probability distribution of each drone type, the category corresponding to the highest probability is selected as the recognition result to obtain the drone signal recognition result.

[0066] Specifically, the set of drone radio signal feature vectors is normalized. Normalization involves converting feature vectors of different dimensions and ranges to the same numerical range, ensuring a relatively balanced contribution to the model. Specific methods include min-max normalization and Z-score normalization. Min-max normalization maps feature values to the interval [0, 1]. This is calculated by subtracting the minimum value from the feature value and dividing the result by the difference between the maximum and minimum values. Z-score normalization transforms the feature value into a distribution with a mean of 0 and a standard deviation of 1. This is calculated by subtracting the mean from the feature value and dividing the result by the standard deviation. In drone signal processing, min-max normalization is typically used for frequency-domain features such as center frequency and bandwidth, while Z-score normalization is used for time-domain features such as signal strength. After normalization, the feature vectors are reorganized into a format suitable for convolutional neural network input. Typically, the one-dimensional feature vectors are reshaped into a two-dimensional matrix or a three-dimensional tensor to facilitate convolution operations.

[0067] The training data is fed into a network structure consisting of five convolutional layers for feature mapping, resulting in hierarchical feature representation data. The convolutional layer is the core component of a convolutional neural network, extracting local features by sliding a convolution kernel over the input data. In the drone signal recognition network, the first convolutional layer uses 32 3×3 convolution kernels with a stride of 1 and padding of 1, primarily extracting low-level features such as edges and textures. The second convolutional layer uses 64 3×3 convolution kernels to extract more complex local structures. The third convolutional layer uses 128 3×3 convolution kernels to capture higher-level feature combinations. The fourth convolutional layer uses 256 3×3 convolution kernels. The fifth convolutional layer uses 512 3×3 convolution kernels to extract highly abstract feature representations. The ReLU activation function is applied after each convolution layer to introduce nonlinearity and enhance the network's expressive power. The ReLU function sets negative values to 0 and retains positive values. Its simple yet effective form helps address the vanishing gradient problem in deep networks. The five convolutional layers extract features step by step to form hierarchical feature expression data, from simple signal features to complex drone feature patterns.

[0068] The hierarchical feature expression data is processed through max pooling and batch normalization layers to reduce dimensionality and normalize, resulting in compressed feature expression data. The max pooling layer is a downsampling operation that selects the maximum value within a fixed-size window as the output, thereby reducing the data dimension. In the drone signal recognition network, each convolutional layer is followed by a 2×2 max pooling layer with a stride of 2, which halves the width and height of the feature map, significantly reducing the amount of subsequent computation. The batch normalization layer normalizes each mini-batch of data, stabilizing its distribution within a range of mean 0 and variance 1, helping to accelerate network training and improve generalization. Batch normalization is performed after each convolutional layer and before the activation function. The process involves calculating the mean and variance of the batch data, then normalizing it. Finally, learnable scaling and translation parameters are introduced to enable the network to adaptively adjust the data distribution. Through the combination of max pooling and batch normalization, the hierarchical feature expression data is compressed into a more compact and standardized representation.

[0069] The compressed feature expression data is input into two fully connected layers for nonlinear transformation, resulting in a vector representation of the data in a high-dimensional feature space. The fully connected layer connects all input neurons to all output neurons, integrating the local features extracted by the previous convolutional layer to form a global feature representation. In the drone signal recognition network, the first fully connected layer contains 1024 neurons, which receive the compressed feature expression data. After performing a linear transformation, it uses the ReLU activation function to introduce nonlinearity. A 50% dropout is applied to randomly inactivate half of the neurons to prevent overfitting. The second fully connected layer contains 512 neurons, which further integrates the features, also using ReLU activation and 50% dropout. The two-layer fully connected network maps the compressed feature expression data into a high-dimensional feature space, forming a more abstract vector representation that is closer to the semantic-level characteristics of the drone type.

[0070] The Softmax function is used to calculate the probability distribution of the vector data in the high-dimensional feature space, yielding the probability distribution of each drone type. The Softmax function converts any real-valued vector into a probability distribution and is commonly used in the output layer of multi-classification problems. It exponentially divides each element in the input vector by the sum of all exponentials, ensuring that the sum of the output elements is 1, which can be interpreted as the class probability. In the drone signal recognition network, the Softmax layer receives the output vector of the second fully connected layer. The number of neurons in the layer equals the number of drone types plus one (one additional class represents non-drone signals). For example, if the system needs to identify five drone types, the Softmax layer has six neurons, corresponding to the probabilities of the five drone types and one non-drone signal. The Softmax calculation converts the vector data into probability values in the range [0, 1] that sum to 1, intuitively reflecting the probability of the signal belonging to each class. Based on the probability distribution of each drone type, the class with the highest probability is selected as the recognition result, resulting in the drone signal recognition result. This step transforms the probability distribution into a decision, employing the maximum a posteriori probability criterion, which selects the class with the highest probability as the recognition result. In practical applications, a probability threshold, such as 0.75, is often set. Recognition is confirmed only when the highest probability exceeds the threshold; otherwise, it is marked as "unknown type" to improve recognition reliability. For recognition of consecutive frames, temporal smoothing strategies, such as majority voting or exponentially weighted averaging, can also be employed to combine multi-frame results for more stable recognition. The output drone signal recognition result includes the identified drone type, recognition confidence, and recognition timestamp, providing the basic input for subsequent joint recognition of clustered signals.

[0071] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0072] A two-dimensional self-organizing mapping network is configured for the drone signal recognition results. The network is trained by setting the batch mode, initial learning rate, learning rate decay function, initial neighborhood radius and neighborhood shrinkage function to obtain a topology-preserving mapping network.

[0073] The known UAV sample features are input into the topology-preserving mapping network for mapping, and the distribution of various samples on the grid is counted to obtain the grid position probability distribution map;

[0074] The drone signal recognition result is input into the topology-preserving mapping network, the distance value with each neuron is calculated, and the neuron with the smallest distance is selected as the matching unit to obtain the position of the signal in the feature space;

[0075] According to the position of the signal in the feature space, the grid position probability distribution map is queried and the spatial mapping probability value is obtained through weighted summation of the adjacent areas.

[0076] The spatial mapping probability value and the classification probability value in the drone signal recognition result are weighted and fused according to the dynamically adjusted weight coefficient to obtain the comprehensive recognition probability;

[0077] The state transition matrix is applied to the comprehensive recognition probability for time series smoothing, and voting statistics are performed based on the historical frame recognition results to obtain the drone type recognition results.

[0078] Specifically, a two-dimensional self-organizing map (SOM) network is configured for drone signal recognition results. The network is trained by setting batch mode, an initial learning rate, a learning rate decay function, an initial neighborhood radius, and a neighborhood shrinkage function. Obtaining a topology-preserving SOM network is a critical first step. A two-dimensional SOM network is an unsupervised artificial neural network that maps high-dimensional input data into a low-dimensional space (usually two-dimensional) while preserving the data's topological structure. In this drone detection method, the SOM network consists of a 10×10 two-dimensional grid with 100 neurons. Each neuron has a weight vector with the same dimensions as the input data. The network is trained in batch mode, processing multiple samples at once and updating cumulatively, which is more stable than single-sample updates. The initial learning rate is set to 0.5 and gradually decreased using an exponential decay function (learning rate = initial learning rate × exp(-current iteration number / total iteration number)). This allows for large-scale adjustments early on and fine-tuning later on. The neighborhood radius is initially set to 5, indicating a range of influence covering five cells around the center point. It decreases linearly to 1 as training progresses, ensuring global sorting early on and local fine-tuning later on. During training, for each input sample, the neuron that best matches it (called the winning neuron) is found. The weight vectors of this neuron and its neighbors are then updated to make them closer to the input sample. After 1,000 rounds of training, the network converges and forms a topology-preserving mapping of the input data space.

[0079] The features of known drone samples are input into a topology-preserving mapping network for mapping. The distribution of each type of sample on the grid is then statistically analyzed to generate a probability distribution map for each grid location. This step establishes an association between categories and grid locations. Known drone samples refer to a dataset of drone signal features whose types have been confirmed, typically collected in the laboratory or from field recordings. These samples are then fed into the trained topology-preserving mapping network. For each sample, the best matching unit in the network is identified. This involves calculating the Euclidean distance between the sample and the weight vector of each neuron, and selecting the neuron with the smallest distance as the activation point. The activation distribution of each type of drone sample on the grid is then statistically analyzed. The frequency of activation of each grid location by each type of sample is calculated, and the frequency is converted into a probability: the number of activations of a sample of a certain type at that location divided by the total number of samples of that type. This results in a three-dimensional data structure: grid coordinate x, grid coordinate y, and category, with the value representing the probability of activation of the corresponding category at that location. This grid location probability distribution map visually illustrates the distribution of different types of drones in feature space, providing a spatial reference for subsequent sample classification.

[0080] The drone signal recognition results are input into a topology-preserving mapping network, where the distance to each neuron is calculated. The neuron with the smallest distance is selected as the matching unit, determining the signal's location in feature space. The drone signal recognition results, which contain feature representations extracted by the convolutional neural network, are then input into the trained topology-preserving mapping network, where the distance to each neuron's weight vector is calculated. Distance is typically calculated using the Euclidean distance, which is the square root of the sum of the squared differences between the corresponding elements of two vectors. After the calculation, the neuron with the smallest distance is selected as the best matching unit, representing the signal's mapping location in the two-dimensional feature space. This location is a two-dimensional coordinate (x, y), where both x and y values range from 0 to 9 (corresponding to a 10×10 grid). In this way, high-dimensional drone signal features are mapped to a specific location in two-dimensional space, facilitating visualization and comparison with the distribution of known samples.

[0081] Based on the signal's position in feature space, the grid location probability distribution map is queried. A weighted summation of the neighboring regions is performed to obtain a spatial mapping probability value. This step infers the signal's likely category based on its spatial location. First, based on the signal's position (x, y) in feature space obtained in the previous step, the probability values for each category at that location in the grid location probability distribution map are queried. Taking into account the continuity and smoothness of the feature map, not only the probability of the exact matching location is queried, but also the probability distribution of neighboring locations is considered, using a weighted summation approach. The specific method is to take all grid points within a neighborhood with a certain radius (usually 2) centered on the signal mapping location and perform a weighted summation of the probability values for each category at each grid point. The weight decreases with distance, typically using a Gaussian weighting function. The result is a probability vector, where each element corresponds to the probability of a drone type, reflecting the likelihood of a category assignment based on spatial location. This neighborhood weighting helps smooth out local fluctuations and improve classification stability.

[0082] The spatial mapping probability value and the classification probability value from the drone signal recognition results are weighted and fused using dynamically adjusted weight coefficients to produce a comprehensive recognition probability. This step combines the classification results of the convolutional neural network and the spatial clustering results of the self-organizing map, which complement each other. The classification probability value in the drone signal recognition results is the probability distribution output by the softmax layer of the convolutional neural network, while the spatial mapping probability value is the category distribution inferred based on the signal's position in feature space. The two probabilities are fused using a weighted average, with the weight coefficient dynamically adjusted based on the signal-to-noise ratio (SNR). When the SNR is high, the convolutional neural network classification result is given a higher weight, for example, 0.7; when the SNR is low, the spatial mapping result is given a higher weight, for example, 0.3. The fusion formula is: Comprehensive Probability = α × Convolutional Neural Network Probability + (1-α) × Spatial Mapping Probability, where α is the dynamically adjusted weight coefficient. The fused comprehensive recognition probability combines the advantages of both methods, leveraging the convolutional neural network's sensitivity to features and the self-organizing map's ability to capture spatial distributions.

[0083] The state transition matrix is applied to the overall recognition probability for time series smoothing. Voting is then combined with historical frame recognition results to generate the drone type recognition result. This step takes into account the temporal continuity of the drone signal and improves recognition stability through time series smoothing. The state transition matrix describes the transition probability from the current frame's drone type to the next frame's type and is typically derived from historical data. Applying the state transition matrix adjusts the overall recognition probability for the current frame based on the previous frame's recognition results and state transition probabilities, strengthening temporally continuous class judgment. Time series smoothing also involves performing a sliding window voting on the recognition results of multiple consecutive frames. The window size is typically 5 to 10 frames, and the most frequently occurring category within the window is selected as the result. This temporal processing effectively reduces transient interference and misjudgments, improving overall recognition stability and reliability. The output drone type recognition result more accurately reflects the actual type of the detected target.

[0084] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0085] The roles of master node, relay node and edge node are assigned to sensor nodes, and communication links between nodes are established through a hierarchical network topology to obtain a multi-level relay network architecture.

[0086] Perform precise time protocol synchronization on each node in the multi-level relay network architecture to keep the clock deviation of each node within the specified range, thus obtaining a time-synchronized node network;

[0087] Perform trilateration calculation on the drone type identification results based on the geographical location and signal reception strength of each node to obtain the signal source location coordinates;

[0088] The feature vectors extracted by each node are transmitted to the upper node through a multi-level relay network architecture, and feature weighted averaging and dimensionality reduction are performed to obtain a fused feature vector;

[0089] The confidence and uncertainty of each node's independent recognition results are calculated using the evidence theory framework, and the weight is dynamically adjusted according to the signal-to-noise ratio and distance to obtain a comprehensive recognition result.

[0090] The signal source position coordinates are correlated and matched with the comprehensive recognition results, and the trajectory is associated through the tracking filter to obtain the drone position and identity information.

[0091] Specifically, assigning sensor nodes the roles of master, relay, and edge nodes, and establishing inter-node communication links through a hierarchical network topology to achieve a multi-level relay network architecture is the first step in building a collaborative detection network. Sensor nodes are radio signal receiving devices distributed within a monitoring area. They are divided into three roles based on their function and location: master nodes are responsible for global data fusion and decision-making, typically equipped with high computing power and storage resources, and are located at the top layer of the network; relay nodes are responsible for data forwarding and regional fusion, located in the middle layer of the network, acting as a communication bridge between nodes at upper and lower layers; edge nodes focus on local signal acquisition and preliminary processing, directly sensing the environment, and are located at the bottom layer of the network. Role assignment is based on the node's hardware performance, geographic location, and coverage, and is determined by evaluating factors such as computing power, storage capacity, communication bandwidth, and power durability. The hierarchical network topology is a tree-like structure, with edge nodes connected to relay nodes, which in turn connect to master nodes, forming a multi-level data aggregation path. Inter-node communication links utilize secure and encrypted wireless communication technologies, such as dedicated encrypted Wi-Fi or low-power wide area network technologies, to ensure the security and reliability of data transmission. Each node in a multi-level relay network architecture is synchronized using the Precision Time Protocol (PTP) to maintain clock deviation within a specified range, resulting in a time-synchronized node network. Time synchronization is fundamental to collaborative detection, as accurate timestamps are crucial for correlating multi-source signals and locating targets. The Precision Time Protocol (PTP) is a high-precision time synchronization method implemented in drone detection networks using the IEEE 1588 Precision Time Protocol (PTP). The synchronization process first identifies a master clock node, typically selected as the time reference. Time information is then synchronized to all nodes at each level through a hierarchical transmission process. Synchronization messages contain send and receive timestamps, as well as propagation delay estimates. Receiving nodes use this information to calculate the deviation between their own clocks and the master clock and make appropriate adjustments. To account for the uncertainty of wireless communications, multiple measurements are averaged to reduce random errors, and an outlier detection mechanism is implemented to eliminate significantly erroneous synchronization data. Through PTP synchronization, the clock deviation of each node is kept within 100 nanoseconds, meeting the accuracy requirements for drone signal correlation and positioning.

[0092] The drone type identification results are trilaterated based on the geographic location and signal reception strength of each node to obtain the signal source location coordinates. Trilateration is a distance-based positioning method that determines the target's spatial position by measuring the distance from the target to three or more nodes with known locations. In radio detection, direct distance measurement is difficult and is often estimated indirectly through signal reception strength. First, according to radio propagation models, signal strength attenuates with distance, following an inverse square law or more complex models that account for path loss, shadowing, and multipath. Received signal strength is related to transmit power, antenna gain, propagation distance, and environmental factors. Through reverse calculation, the distance from the transmitting source to the receiving point is inferred based on the received signal strength. When three or more nodes simultaneously receive the same signal and calculate their respective estimated distances, the signal source's location can be determined by solving a system of geometric equations. This calculation utilizes a weighted least squares method, with weights proportional to the signal strength and the node's historical positioning accuracy to minimize the impact of measurement errors. The resulting drone location coordinates include longitude, latitude, and altitude, and their accuracy is closely related to node deployment density and signal conditions.

[0093] The feature vectors extracted by each node are transmitted to the upper-level node via a multi-level relay network architecture. Weighted averaging and dimensionality reduction are performed to produce a fused feature vector. The feature vector is the drone's characteristic extracted by each node based on the locally received signal. It includes frequency, time, modulation, and protocol features. The feature vector is uploaded via the multi-level relay network. The edge node sends the feature vector to its connected relay node. The relay node collects the feature vectors from the lower-level nodes, performs preliminary fusion, and then uploads them to the master node. Weighted averaging is a feature-level fusion method that calculates a weighted average of similar features provided by different nodes. The weight coefficient is determined based on the node's signal-to-noise ratio, signal integrity, and historical recognition accuracy. Nodes with high signal-to-noise ratio, high signal integrity, and high historical accuracy are given higher weights. Feature dimensionality reduction is the process of extracting key information from a high-dimensional feature space. Methods such as principal component analysis (PCA) and linear discriminant analysis (LDA) are commonly used to reduce the data dimension while preserving the variation in key features, improving subsequent processing efficiency. The fused feature vector integrates observational information from multiple nodes and more comprehensively reflects the characteristics of the drone than single-node features.

[0094] The confidence and uncertainty of the independent identification results of each node are calculated through the evidence theory framework, and the weights are dynamically adjusted according to the signal-to-noise ratio and distance to obtain a comprehensive identification result. The evidence theory framework is a method for processing uncertainty and fusing multi-source information. It is particularly suitable for processing the judgment results of different nodes on the same target. In drone identification, each node obtains the judgment result of the drone type based on local observations, including category judgment and confidence. In evidence theory, the basic probability distribution function (BPA) is used to represent the node's support for each category, the confidence represents the direct support for a certain category, and the uncertainty represents the reserved judgment when there is insufficient evidence. The BPA function is defined as follows:

[0095] ;

[0096] in, is the category of node i The basic probability distribution of is the signal-to-noise ratio of node i, is a node distance to the target, is the distance influence factor, Node i determines whether the target belongs to the category The denominator is the normalized term of the weighted sum of all nodes, N is the total number of nodes involved in the fusion, is a node Distance to target.

[0097] When multiple nodes have different judgments on the same target, the BPA of each node is fused through the Dempster combination rule to obtain the comprehensive BPA:

[0098] ;

[0099] here and They are the BPA of different nodes, and the summation range is all intersections The summation range in the denominator is the combination of all empty sets, indicating the degree of conflict. In this way, the judgments of each node are comprehensively considered to obtain more reliable drone type identification results.

[0100] The signal source's location coordinates are correlated with the comprehensive recognition results, and the trajectory is correlated using a tracking filter to obtain the drone's location and identity information. Correlation matching is the process of binding the positioning and recognition results to ensure that the location and identity information correspond to the same target. Correlation matching becomes particularly important when multiple targets are present in the environment. The fundamental principle of correlation is spatiotemporal consistency: signals at similar locations within the same time window are likely to originate from the same target. A tracking filter is a recursive estimation method used to extract the target's true state from noisy observation data. Kalman filters or particle filters are commonly used in drone detection. The tracking process first establishes a drone motion model that describes the dynamic changes in state variables such as position, velocity, and acceleration. New observations are then compared with the model's predicted values to update the state estimate. Tracking filtering not only smooths observation noise but also provides a prediction of the target's motion trends, facilitating early identification of potential threats. Through continuous tracking, the drone's trajectory is generated. Combined with its identity information, this provides a comprehensive description of the target's activity characteristics, providing a basis for subsequent threat assessment and defense decisions.

[0101] In a drone detection experiment, a three-level relay network architecture was deployed, consisting of one master node, three relay nodes, and eight edge nodes. Time synchronization among all nodes was achieved using the IEEE 1588 Precision Time Protocol, with clock deviations kept within 75 nanoseconds. When a drone entered the monitoring area, five edge nodes simultaneously received signals with signal strengths of -65dBm, -68dBm, -72dBm, -75dBm, and -80dBm, respectively. The distances from the nodes to the drone were estimated using radio propagation models, and the drone's position coordinates were calculated using trilateration. Simultaneously, the feature vectors extracted by the five nodes were uploaded to the master node via the relay nodes. After weighted averaging and PCA dimensionality reduction, a fused feature vector was generated. Each node independently identified the drone type: three nodes identified it as Model A with confidence levels of 0.85, 0.78, and 0.72, respectively; one node identified it as Model B with a confidence level of 0.65; and one node identified it as Model C with a confidence level of 0.60. Applying the evidence theory framework, the BPA was calculated and fused, taking into account the signal-to-noise ratio (SNR) of each node (22dB, 20dB, 18dB, 15dB, and 12dB, respectively) and the distance. The final comprehensive recognition result was model A with a confidence level of 0.89. The position coordinates were associated with the recognition result, and the trajectory was tracked using a Kalman filter to obtain the drone's location and identity information.

[0102] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0103] The threat index is calculated based on the drone's location and identity information, combined with the drone's type risk factor, proximity, size, approach speed, and regional safety level.

[0104] Thresholds are set for the threat index, and the threat level is divided into four levels: normal, caution, warning, and emergency, to obtain the warning level determination result;

[0105] Adjust the scanning frequency of the monitoring area based on the warning level determination results, and send different priority notifications to relevant personnel to obtain preliminary response measures;

[0106] Based on the warning level determination result, the corresponding detection equipment is activated through the multi-level relay network to confirm the target and obtain target confirmation information;

[0107] Track the trajectory of the drone and predict its position and intention through Kalman filtering to obtain the trajectory prediction result;

[0108] The target confirmation information and trajectory prediction results are combined with the warning level determination results to allocate response resources and obtain the defense response strategy.

[0109] Specifically, a key step in achieving intelligent early warning is to calculate a threat index based on drone location and identity information, combined with the drone type's risk factor, proximity, size, closing speed, and regional security level. The threat index is a quantitative measure of the threat posed by drones to a specific area, calculated by comprehensively considering multiple factors. The drone type's risk factor is a predefined weighted value based on the drone's model, purpose, and potential hazard. For example, a military drone has a higher risk factor than a civilian drone, and a drone equipped with a camera has a higher risk factor than an ordinary toy drone. Proximity refers to the minimum distance between a drone and the boundary of a sensitive area; the closer the distance, the greater the threat. Drone size refers to the drone's physical dimensions or weight. Larger drones generally have greater payload capacity and potential destructive power. Closing speed refers to the component of a drone's velocity toward a sensitive area. Faster speeds reduce the time available for response systems and increase the threat. Regional security level is a predefined measure of the importance of different areas. For example, critical infrastructure such as nuclear facilities and government buildings has a higher security level. These factors are weighted to create a comprehensive threat index using the following formula:

[0110] ;

[0111] in, Threat Index, the value range is [0,1]; Indicates the drone category risk factor (Drone Category), which is usually preset in the range of [0,1]; Indicates the proximity (ApproachProximity), by converting the actual distance into a normalized value in the [0,1] interval. The closer the distance, the larger the value. Indicates the size of the drone (Size), which is also normalized to the interval [0,1]; Represents the approach speed (Speed Vector), normalized to the [0,1] interval; Indicates the security level of the area (Security Area), which is preset in the range [0,1]. Weight coefficient 、 、 、 and Corresponding to the importance weight of each factor, and satisfying + + + + = 1, ensuring the threat index is between 0 and 1.

[0112] Thresholds are applied to the threat index, categorizing the threat level into four levels: normal, caution, warning, and emergency, to generate the warning level determination result. Thresholding is the process of mapping continuous threat index values to discrete warning levels. Based on practical application experience and security requirements, the threat index range of 0 to 1 is divided into four sub-ranges: 0 to 0.25 is the normal level, indicating no obvious threat; 0.25 to 0.5 is the caution level, indicating a potential threat requiring attention; 0.5 to 0.75 is the warning level, indicating a moderate threat requiring defensive measures; and 0.75 to 1.0 is the emergency level, indicating a severe threat requiring immediate action. Thresholding is not a simple uniform division; rather, it is adjusted based on actual security requirements and response capabilities. For example, in scenarios with high security requirements, the emergency level threshold can be lowered to trigger a higher-level warning earlier. The warning level determination result is a discrete value that is directly linked to the subsequent response strategy.

[0113] Based on the alert level, the scanning frequency of the monitored area is adjusted, and notifications of varying priorities are sent to relevant personnel to initiate initial response measures. Scanning frequency adjustment dynamically allocates detection resources based on threat level, with higher-threat areas receiving higher scanning frequencies to improve detection accuracy and timeliness. Specifically, the following adjustments are implemented: Normal-level areas maintain a basic scanning frequency, such as once every 10 seconds; Caution-level areas double the scanning frequency, such as once every 5 seconds; Warning-level areas quadruple the scanning frequency, such as once every 2.5 seconds; and Emergency-level areas increase the scanning frequency to the maximum, such as once every 1 second. Notification delivery is the process of communicating alert information to relevant personnel, with different levels corresponding to different notification methods and recipients. The Normal-level is typically recorded only in the system log, with no proactive notifications. The Caution-level sends low-priority notifications to monitoring operators, such as system interface prompts. The Warning-level sends high-priority notifications, such as text messages and emails, to security supervisors and response personnel. The Emergency-level sends urgent notifications, such as phone calls and sirens, to all security personnel and management, ensuring immediate attention. Initial response measures include resource preparation and initial defensive actions, laying the foundation for a more in-depth response.

[0114] Based on the warning level determination, a multi-level relay network activates the corresponding detection equipment for target confirmation, generating target confirmation information. Target confirmation cross-validates radio detection results, providing supplementary evidence by activating other types of detection equipment. Different warning levels trigger different levels of confirmation mechanisms: the caution level may only activate the radio sensing functions of adjacent nodes; the warning level activates more radio nodes and also optoelectronic equipment such as visible light and infrared cameras for visual confirmation; the emergency level activates all available resources, including high-precision radar and acoustic detectors. The multi-level relay network plays a role in command transmission and data collection in this process, issuing activation commands from the master node to the corresponding level of detection equipment and aggregating detection results up the network. Target confirmation information includes the fusion results of multiple sensors, such as target outline features in optoelectronic images and radar echo characteristics, combined with radio characteristics to significantly improve the reliability of the judgment.

[0115] The UAV's position information is tracked, and its position and intention are predicted using a Kalman filter to produce a trajectory prediction. Trajectory tracking continuously records the UAV's position changes, generating time-series position data. The Kalman filter is a recursive estimation algorithm that optimally estimates observational data containing random errors and is widely used in target tracking. The Kalman filter consists of two main steps: prediction and update. The prediction step predicts the current state based on the previous state and motion model; the update step corrects the prediction based on new observations. In UAV trajectory tracking, state variables typically include position, velocity, and acceleration. The motion model is determined based on the UAV's flight characteristics, such as a constant velocity model or a constant acceleration model. Position prediction calculates the UAV's probable position at a certain point in the future, based on its current state and motion trends. Intent prediction infers the UAV's intended behavior, such as whether it is passing by, hovering for observation, or preparing to invade. This is achieved by analyzing trajectory patterns and regional relationships. Trajectory prediction results include predicted position coordinates, velocity vectors, and possible behavioral intentions, providing forward-looking information for response decisions.

[0116] Target confirmation information, trajectory prediction results, and warning level determination results are combined to allocate response resources and develop a defense response strategy. Response resource allocation is the process of rationally dispatching limited defense resources based on the threat situation, including passive monitoring resources and active defense equipment. Target confirmation information provides definitive evidence of the current threat, trajectory prediction results provide threat development trends, and warning level determination results provide threat severity. These three elements combine to form a comprehensive situational awareness framework, guiding resource allocation decisions. The specific allocation strategy varies depending on the warning level: the Normal level allocates only basic monitoring resources; the Attention level increases monitoring density and prepares some defense equipment; the Warning level activates most defense equipment and enters a state of readiness; and the Emergency level fully deploys all defense resources, entering a state of maximum alert. A defense response strategy is a detailed action plan that includes the defensive measures to be taken, the order of implementation, responsible personnel, and expected results, ensuring coordinated action against drone threats.

[0117] The above describes the drone detection method based on radio intelligent fusion in the embodiment of the present application. The following describes the drone detection system based on radio intelligent fusion in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the UAV detection system based on radio intelligent fusion includes:

[0118] The acquisition module is used to collect radio signals through distributed relay sensor nodes to obtain digital I / Q data;

[0119] An extraction module, configured to perform time-frequency domain conversion and feature extraction on the digitized I / Q data to obtain a set of UAV radio signal feature vectors;

[0120] An input module, configured to input the drone radio signal feature vector set into a convolutional neural network for training and recognition, thereby obtaining a drone signal recognition result;

[0121] an identification module, configured to identify the drone signal using a clustering signal joint identification algorithm to obtain a drone type identification result;

[0122] A fusion module is used to construct a multi-level relay network based on the UAV type identification results to perform collaborative data fusion processing to obtain the UAV location and identity information;

[0123] The classification module is used to calculate the threat index of the drone's location and identity information and perform early warning classification based on the relay network to obtain a defense response strategy.

[0124] Through the collaborative cooperation of the above components, the radio signals are collected by distributed relay sensor nodes to obtain digital I / Q data, which realizes seamless coverage of wide-area space and solves the problem of limited coverage of a single sensor. The digital I / Q data is converted into time and frequency domains and features are extracted to obtain the feature vector set of UAV radio signals, which fully explores the multi-dimensional features of UAV signals and improves the richness and discrimination of feature expression. The feature vector set of UAV radio signals is input into the convolutional neural network for training and recognition to obtain the UAV signal recognition result, so that the model can automatically learn the deep feature pattern of the signal, and the recognition ability is no longer limited by the manually preset feature rules. The clustering signal joint recognition algorithm is used to The preliminary drone signal recognition results are processed to obtain the drone type recognition results. The cluster analysis and pattern recognition technologies are cleverly integrated, and drone models with similar characteristics can be accurately distinguished, significantly reducing the misidentification rate. According to the drone type recognition results, a multi-level relay network is constructed to perform collaborative data fusion processing to obtain the drone location and identity information, solving the information island problem in traditional single-node detection and realizing effective complementarity and redundant verification of information between different nodes. The threat index is calculated for the drone location and identity information and an early warning classification based on the relay network is performed to obtain the defense response strategy, building a complete intelligent defense chain from perception to decision-making, ensuring that corresponding defense measures are taken against drones of different threat levels. The present invention applies artificial intelligence algorithms and models in specific drone detection application fields, especially convolutional neural networks and clustering signal joint recognition algorithms. It not only surpasses traditional methods in the feature extraction and recognition stages, but also significantly improves detection accuracy and anti-interference ability through algorithm features, and shows excellent adaptability in complex electromagnetic environments; the collaborative data fusion mechanism of the multi-level relay network fully considers the spatial characteristics of radio signal propagation, and utilizes the spatial diversity formed by distributed sensing nodes to maintain stable and reliable detection effects in strong interference environments; the adaptive early warning classification strategy based on threat index deeply combines artificial intelligence decision-making technology with professional field knowledge to form an intelligent defense system against drone threats. Compared with traditional static defense strategies, it has higher targeting and resource utilization efficiency.

[0125] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3As shown. The computer device includes a processor, memory, display screen, input device, network interface and database connected via a system bus. The processor of the computer design is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, the above method is implemented.

[0126] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0127] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which implements the above-described method when executed by a processor. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0128] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware using a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media provided herein and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM.

[0129] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0130] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0131] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

[0132] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which implements the above-described method when executed by a processor. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0133] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A UAV detection method based on radio intelligent fusion, characterized in that: The UAV detection method based on radio intelligent fusion includes: Radio signals are collected through distributed relay sensor nodes to obtain digital I / Q data; Performing time-frequency domain conversion and feature extraction on the digitized I / Q data to obtain a UAV radio signal feature vector set; Inputting the drone radio signal feature vector set into a convolutional neural network for training and recognition to obtain a drone signal recognition result; The UAV signal recognition result is processed by a clustering signal joint recognition algorithm to obtain a UAV type recognition result, including: configuring a two-dimensional self-organizing mapping network for the UAV signal recognition result, and performing network training by setting a batch mode, an initial learning rate, a learning rate decay function, an initial neighborhood radius, and a neighborhood shrinkage function to obtain a topology-preserving mapping network; inputting known UAV sample features into the topology-preserving mapping network for mapping, and statistically analyzing the distribution of various types of samples on the grid to obtain a grid position probability distribution map; inputting the UAV signal recognition result into the topology-preserving mapping network, calculating the distance value with each neuron, selecting the neuron with the smallest distance as the matching unit, and obtaining the position of the signal in the feature space; querying the grid position probability distribution map according to the position of the signal in the feature space, and obtaining a spatial mapping probability value through weighted summation of neighboring areas; weightedly fusing the spatial mapping probability value with the classification probability value in the UAV signal recognition result according to a dynamically adjusted weight coefficient to obtain a comprehensive recognition probability; applying a state transition matrix to the comprehensive recognition probability for time series smoothing, and performing voting statistics in combination with historical frame recognition results to obtain the UAV type recognition result; Constructing a multi-level relay network based on the drone type identification results to perform collaborative data fusion processing to obtain the drone location and identity information; The threat index is calculated based on the position and identity information of the UAV and a warning classification based on the relay network is performed to obtain a defense response strategy.

2. The UAV detection method based on radio intelligent fusion according to claim 1 is characterized in that: The method of collecting radio signals through distributed relay sensor nodes to obtain digital I / Q data includes: The radio signal in the environment is pre-amplified by a low-noise amplifier to obtain an analog radio frequency signal; filtering out-of-band interference on the analog radio frequency signal through a bandpass filter to obtain a filtered radio frequency signal; Down-converting the filtered radio frequency signal using a mixer and a local oscillator to obtain an intermediate frequency or baseband signal; Digitizing the intermediate frequency or baseband signal at a high sampling rate through an analog-to-digital converter to obtain an original I / Q data stream; Performing digital downsampling and digital filtering on the original I / Q data stream to obtain noise-reduced I / Q data; The noise reduction I / Q data is marked by adding a timestamp and geographic location information to obtain the digitized I / Q data.

3. The UAV detection method based on radio intelligent fusion according to claim 1 is characterized in that: The time-frequency domain conversion and feature extraction of the digitized I / Q data to obtain a UAV radio signal feature vector set includes: Performing time-frequency domain conversion on the digitized I / Q data by short-time Fourier transform to obtain a signal spectrogram; Performing power spectrum density calculation and frequency occupied bandwidth analysis on the signal spectrum to obtain frequency domain characteristic parameters; Extracting time domain features from the digitized I / Q data through envelope analysis to obtain time domain features; Performing high-order cumulants and cyclic spectrum feature calculations on the time domain features and the frequency domain feature parameters to obtain modulation recognition parameters; Performing MAC layer feature extraction on the modulation identification parameters and signal message structure to obtain communication protocol features; The frequency domain feature parameters, the time domain features, the modulation identification parameters and the communication protocol features are normalized and weighted to obtain the UAV radio signal feature vector set.

4. The UAV detection method based on radio intelligent fusion according to claim 1 is characterized in that: The step of inputting the UAV radio signal feature vector set into a convolutional neural network for training and recognition to obtain a UAV signal recognition result includes: Normalizing the UAV radio signal feature vector set to obtain training data that conforms to a convolutional neural network input format; The training data is fed into a network structure comprising five convolutional layers for feature mapping to obtain hierarchical feature expression data; Performing dimensionality reduction and standardization processing on the hierarchical feature expression data through a maximum pooling layer and a batch normalization layer to obtain compressed feature expression data; Inputting the compressed feature expression data into two fully connected layers for nonlinear transformation to obtain vector representation data in a high-dimensional feature space; Performing probability distribution calculation on the vector representation data in the high-dimensional feature space using the Softmax function to obtain the probability distribution of each drone type; Based on the probability distribution of each drone type, the category corresponding to the highest probability is selected as the recognition result to obtain the drone signal recognition result.

5. The UAV detection method based on radio intelligent fusion according to claim 1 is characterized in that: The multi-level relay network is constructed based on the drone type identification result to perform collaborative data fusion processing to obtain the drone location and identity information, including: The roles of master node, relay node and edge node are assigned to sensor nodes, and communication links between nodes are established through a hierarchical network topology to obtain a multi-level relay network architecture. Performing precise time protocol synchronization on each node in the multi-stage relay network architecture to keep the clock deviation of each node within a specified range, thereby obtaining a time-synchronized node network; Perform trilateration calculation on the drone type identification result according to the geographical location and signal reception strength of each node to obtain the signal source location coordinates; The feature vectors extracted by each node are transmitted to the upper node through the multi-level relay network architecture, and feature weighted averaging and dimensionality reduction processing are performed to obtain a fused feature vector; The confidence and uncertainty of each node's independent recognition results are calculated using the evidence theory framework, and the weight is dynamically adjusted according to the signal-to-noise ratio and distance to obtain a comprehensive recognition result. The signal source position coordinates are associated and matched with the comprehensive recognition results, and trajectory association is performed through a tracking filter to obtain the drone position and identity information.

6. The UAV detection method based on radio intelligent fusion according to claim 1 is characterized in that: The calculation of the threat index based on the position and identity information of the drone and the execution of early warning classification based on the relay network to obtain a defense response strategy include: A threat index is obtained by performing a weighted calculation based on the drone's location and identity information, combined with the drone's type risk factor, proximity, size, approach speed, and regional safety level; Threshold classification is performed on the threat index to divide the threat level into four levels: normal, caution, warning, and emergency, and obtain a warning level determination result; Adjust the scanning frequency of the monitoring area according to the warning level determination result, and send different priority notifications to relevant personnel to obtain preliminary response measures; Based on the warning level determination result, corresponding detection equipment is activated through a multi-level relay network to perform target confirmation and obtain target confirmation information; Tracking the position of the UAV, performing position and intention prediction using Kalman filtering, and obtaining a trajectory prediction result; The target confirmation information and the trajectory prediction result are combined with the warning level determination result to perform response resource allocation to obtain the defense response strategy.

7. A UAV detection system based on radio intelligent fusion, used to implement the UAV detection method based on radio intelligent fusion as described in any one of claims 1 to 6, characterized in that: The UAV detection system based on radio intelligent fusion includes: The acquisition module is used to collect radio signals through distributed relay sensor nodes to obtain digital I / Q data; An extraction module, configured to perform time-frequency domain conversion and feature extraction on the digitized I / Q data to obtain a set of UAV radio signal feature vectors; An input module, configured to input the drone radio signal feature vector set into a convolutional neural network for training and recognition, thereby obtaining a drone signal recognition result; an identification module for performing a clustering signal joint identification algorithm on the drone signal identification result to obtain a drone type identification result, including: configuring a two-dimensional self-organizing mapping network for the drone signal identification result, and performing network training by setting a batch mode, an initial learning rate, a learning rate decay function, an initial neighborhood radius, and a neighborhood shrinkage function to obtain a topology-preserving mapping network; inputting known drone sample features into the topology-preserving mapping network for mapping, and statistically analyzing the distribution of various types of samples on the grid to obtain a grid position probability distribution map; inputting the drone signal identification result into the topology-preserving mapping network, calculating the distance value with each neuron, selecting the neuron with the smallest distance as a matching unit, and obtaining the position of the signal in the feature space; querying the grid position probability distribution map based on the position of the signal in the feature space, and obtaining a spatial mapping probability value through weighted summation of neighboring regions; weightedly fusing the spatial mapping probability value with the classification probability value in the drone signal identification result according to a dynamically adjusted weight coefficient to obtain a comprehensive identification probability; applying a state transition matrix to perform time series smoothing on the comprehensive identification probability, and performing voting statistics in combination with historical frame identification results to obtain the drone type identification result; A fusion module is used to construct a multi-level relay network based on the UAV type identification results to perform collaborative data fusion processing to obtain the UAV location and identity information; The classification module is used to calculate the threat index of the drone's location and identity information and perform early warning classification based on the relay network to obtain a defense response strategy.

8. A computer device, characterized in that: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, it implements the drone detection method based on radio intelligent fusion according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that A computer program is stored thereon, and when the computer program is executed by a processor, the processor enables the processor to execute the drone detection method based on radio intelligent fusion as described in any one of claims 1 to 6.

Images