A deep learning-based full-band unmanned aerial vehicle identification regulation system and method

Abstract

The application relates to the technical field of unmanned aerial vehicle management and control, and discloses a full-frequency-band unmanned aerial vehicle identification management system and method based on deep learning; the system comprises a signal processing module, which converts original analog signals into digital signals; a time-frequency graph construction module, which splices time-frequency subgraphs into a time-frequency graph in sequence; a deep learning module, which identifies unmanned aerial vehicles as local unmanned aerial vehicles or out-of-domain unmanned aerial vehicles; a first instruction formulating module, which formulates continuous flight instructions or flight adjustment instructions; and a second instruction formulating module, which formulates sensitive driving instructions or electromagnetic shielding instructions; the application can construct an extraction system for autonomously and hierarchically identifying high complexity features of unmanned aerial vehicles, can satisfy the real-time position monitoring and dynamic management effect of different types of unmanned aerial vehicles in a target management airspace, and can perform targeted signal intervention or forced management on the dynamic flight conditions of local unmanned aerial vehicles and out-of-domain unmanned aerial vehicles, thereby reducing the negative influence of unmanned aerial vehicles on the target management airspace.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control technology, and more specifically, to a deep learning-based full-band UAV identification and control system and method. Background Technology

[0002] In recent years, with the explosive growth of the drone industry, drones have begun to be widely used in many fields such as logistics and distribution, agricultural plant protection, and air traffic control. At the same time, drone signal recognition and automated control technology has also emerged and is constantly being optimized and evolved. This technology can accurately identify the signals of drones that have entered sensitive areas without authorization and take corresponding control measures based on the identification results, thereby ensuring the safety of airspace.

[0003] The patent application with publication number CN120544425A discloses a method and system for UAV signal detection and automated control platform in complex environments. It integrates electromagnetic, optical and acoustic signals through multimodal signal acquisition equipment, introduces micro-meteorological sensors, and combines quantum encrypted communication links to ensure secure and reliable signal transmission. This greatly improves the comprehensiveness and accuracy of UAV signal detection in complex environments and effectively solves the shortcomings of existing technologies in complex environments.

[0004] Existing drone identification and control systems typically employ narrowband scanning or fixed-frequency monitoring signal identification strategies. This makes it difficult to achieve full-band, blind-spot-free signal coverage and capture of the collected drone signals. Consequently, the extracted features from the drone signals are limited in scope and have limited features for identification, reducing the accuracy of identifying different types of drone signals. Furthermore, the uniform and crude interference control approach for drones fails to allow for targeted control operations based on the real-time dynamic position of different drone types, thus failing to achieve differentiated control of different drone types and ultimately reducing the effectiveness of drone identification and control.

[0005] In view of this, the present invention proposes a deep learning-based full-band UAV identification and control system and method to solve the above problems. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a deep learning-based full-band unmanned aerial vehicle (UAV) identification and control system, applied to an air traffic control platform, comprising:

[0007] The signal processing module, with the signal receiving device as the airspace center, determines the target control airspace, collects the raw analog signals in the full frequency band of the target control airspace, and processes and converts the raw analog signals into digital signals.

[0008] The time-frequency diagram construction module divides the digital signal into sub-signal sets labeled with sequence numbers, performs time-frequency analysis on the sub-signal sets, simulates the time-frequency sub-graphs corresponding to the sub-signal sets, and splices the time-frequency sub-graphs into a time-frequency diagram in sequence based on the sequence numbers.

[0009] The deep learning module extracts multi-dimensional features from the time-frequency graph through a deep learning model. These multi-dimensional features include waveform features, modulation mode features, and spectrum features. The module also performs feature attribute analysis on these multi-dimensional features to identify the drone as either a local drone or an external drone.

[0010] The first command formulation module collects the flight speed, flight heading and flight coordinates of the local UAV, formulates the flight path of the local UAV, compares the flight path with the planned path, and formulates the continuous flight command or flight adjustment command.

[0011] The second instruction formulation module divides the target control airspace into sensitive airspace and no-fly airspace, determines the positional relationship between the UAV outside the airspace and the sensitive airspace and no-fly airspace in real time, and formulates sensitive driving-away instructions or electromagnetic shielding instructions.

[0012] Furthermore, the steps for determining the target control area are as follows:

[0013] The real-time power of the signal receiving device is detected by a power sensor, and the power compliance rate is calculated by dividing the real-time power by the rated power.

[0014] Determine the weather and terrain characteristics of the signal receiving equipment at the current moment, summarize the weather characteristics, terrain characteristics, and power compliance rate into real-time factors, and look up the real-time receiving distance of the signal receiving equipment under the real-time factors through the factor comparison table;

[0015] The location coordinates of the signal receiving device are obtained by querying the BeiDou positioning system. The location coordinates are taken as the airspace center and the real-time receiving distance is taken as the radiation standard. Points that are one radiation standard away from the airspace center are recorded as boundary points, and F boundary points are obtained.

[0016] Connect any three boundary points that are not in the same direction to form a triangular grid. After summing up all the triangular grids, they are pieced together to form the airspace boundary, and the area within the airspace boundary is recorded as the target control area.

[0017] Furthermore, the digital signal conversion steps are as follows:

[0018] Adjust the operating frequency band of the receiving antenna on the signal receiving equipment to fully cover the A communication frequency bands in the communication field, collect the electromagnetic wave signals of the target control airspace in the A communication frequency bands, and record them as the original analog signals;

[0019] The frequency range of the bandpass filter is set to be consistent with A communication frequency bands. The original analog signal is filtered by the bandpass filter, and the filtered original analog signal is linearly amplified by a low-noise amplifier to obtain a high-frequency radio frequency signal.

[0020] The high-frequency radio frequency signal is input into the mixer to be down-frequency, and a low-frequency radio frequency signal is output.

[0021] The sampling clock update duration is used as the sampling duration of the analog-to-digital converter. The voltage values ​​of the low-frequency radio frequency signal are sampled at intervals of one sampling duration, and all voltage values ​​are rounded to the nearest discrete level to output a digital signal.

[0022] Furthermore, the steps for partitioning the sub-signal set are as follows:

[0023] The digital signal obtained by rounding is denoted as the first signal, and the digital signal obtained by rounding is denoted as the second signal. The number of the first signal and the number of the second signal are counted respectively, and denoted as the first value and the second value.

[0024] The signal difference is generated by taking the absolute value of the difference between the first and second values. The signal difference is then compared with the number of digital signals to calculate the unit spread ratio.

[0025] The sampling duration is increased by one unit of amplitude to generate the set duration;

[0026] Mark the reception times of all digital signals, and divide the time interval between the first and last reception times into B sets;

[0027] The digital signals received within the set time period are sequentially aggregated to generate B sub-signal sets, and each sub-signal set is labeled with a sequence number starting with 1.

[0028] Furthermore, the simulation method for the time-frequency subgraph is as follows:

[0029] According to the order of reception time, the digital signal of the sub-signal set is divided into C frame time periods, and the last moment of the previous frame time period is adjusted to overlap with the first moment of the next frame time period.

[0030] The digital signals for each of the C frame time periods are multiplied by a window function to form a window, and the windowed digital signals are then subjected to a Fast Fourier Transform sequentially to output amplitude and phase information.

[0031] After summing the amplitude and phase information, a complex spectrum is generated, and modulo and squaring operations are performed on the complex spectrum of C frame time periods to convert the C complex spectrum into C power spectra.

[0032] Using frequency index as rows, time frame index as columns, and power value as element value, arrange the power spectra of C frame time periods into a two-dimensional matrix, and transform the two-dimensional matrix graph to obtain time-frequency sub-graphs. After traversing B sub-signal sets in ascending order of sequence number, simulate B time-frequency sub-graphs.

[0033] Further steps for identifying local or external drones are as follows:

[0034] When the waveform characteristics are consistent with the calibrated waveform characteristics, the fluctuation characteristics are recorded as reported characteristics;

[0035] When the modulation mode characteristics match the calibrated modulation mode characteristics, the modulation mode characteristics are recorded as reported characteristics;

[0036] When the spectral characteristics match the calibrated spectral characteristics, the spectral characteristics are recorded as reported characteristics;

[0037] The number of reported features of the drone is counted. When the number of reported features is 3, the drone is recorded as a local drone.

[0038] When the number of reported features is not 3, the drone will be recorded as an off-domain drone.

[0039] Furthermore, the steps for determining the flight path are as follows:

[0040] The moment when the local drone first enters the target controlled airspace is recorded as the start time, and the time period from the start time to the current time is recorded as the flight period.

[0041] The flight coordinates of the local drone at D times during the flight period are retrieved one by one using the BeiDou positioning system. The points where the D flight coordinates are located are marked one by one on the satellite map. The D points are then connected in sequence to generate the initial path.

[0042] The flight speed and heading of the local drone at D times are queried one by one. After binding the flight speed and heading at the same time, D flight parameters are generated. The D flight parameters are noted on D points respectively, so that the initial path is converted into a flight path.

[0043] Furthermore, the steps for formulating continuous flight instructions or flight adjustment instructions are as follows:

[0044] The planned path of the local drone is retrieved through the air traffic control platform. The part of the planned path corresponding to the flight path is recorded as the target path, and E trajectory points on the target path are marked.

[0045] Find the position coordinates of the points on the flight path that correspond to the E trajectory points one by one, record the trajectory points that have the same position coordinates as the points as coincidence points, count the number of coincidence points, divide the number of coincidence points by the number of trajectory points, and calculate the overlap ratio.

[0046] When the overlap ratio is greater than or equal to the preset overlap threshold, a continuous flight command is issued.

[0047] When the overlap ratio is less than the preset overlap threshold, a flight adjustment command is generated.

[0048] Furthermore, the steps for formulating sensitive expulsion instructions or electromagnetic shielding instructions are as follows:

[0049] The location coordinates of all points in sensitive areas and no-fly zones were queried one by one using the BeiDou positioning system. The location coordinates of sensitive areas and no-fly zones were then summarized to generate sensitive coordinate databases and no-fly coordinate databases.

[0050] The system retrieves the flight coordinates of an out-of-domain drone at the current moment. If the flight coordinates match the location coordinates in the sensitive coordinate database, the out-of-domain drone is located within the sensitive airspace at the current moment, and a sensitive expulsion command is issued.

[0051] When the flight coordinates match the position coordinates in the no-fly zone coordinate database, the UAV outside the zone is located inside the no-fly zone at the current moment, and an electromagnetic shielding command is issued.

[0052] A deep learning-based full-band UAV identification and control method is applied to an air traffic control platform. It is implemented based on a deep learning-based full-band UAV identification and control system, comprising:

[0053] S01: Using the signal receiving device as the airspace center, determine the target control airspace, collect the raw analog signals of the entire frequency band in the target control airspace, and process and convert the raw analog signals into digital signals;

[0054] S02: Divide the digital signal into sub-signal sets labeled with sequence numbers, perform time-frequency analysis on the sub-signal sets, simulate the time-frequency sub-graphs corresponding to the sub-signal sets, and stitch the time-frequency sub-graphs into a time-frequency graph in sequence based on the sequence numbers;

[0055] S03: Extract multi-dimensional features from the time-frequency graph using a deep learning model, and perform feature attribute analysis on the multi-dimensional features to identify the drone as a local drone or an external drone; if it is a local drone, proceed to S04; if it is an external drone, proceed to S05.

[0056] S04: Collect the flight speed, flight heading and flight coordinates of the local UAV, formulate the flight path of the local UAV, compare the flight path with the planned path, and formulate continuous flight command or flight adjustment command.

[0057] S05: Divide the target control airspace into sensitive airspace and no-fly airspace, determine the positional relationship between the UAV outside the airspace and the sensitive airspace and no-fly airspace in real time, and formulate sensitive driving-away instructions or electromagnetic shielding instructions.

[0058] The technical effects and advantages of the deep learning-based full-band drone identification and control system and method of this invention are as follows:

[0059] (1): This invention collects electromagnetic wave signals across the entire frequency band within the target control airspace, enabling comprehensive and seamless collection of electromagnetic wave signals of different frequencies that intrude into the target control airspace. This avoids the omission of collecting electromagnetic wave signals from UAVs with excessively large or small frequencies, thereby achieving comprehensive coverage of UAV signals and ensuring that all electromagnetic wave signals intruding into the target control airspace can be collected, identified, and processed.

[0060] (2): This invention can convert massive discrete electromagnetic wave signals into time-frequency graphs by performing preprocessing and time-frequency analysis on the original analog signals. This allows for the graphical representation of the power intensity of electromagnetic wave signals at different times and frequencies. Combined with a deep learning model, the time-frequency graphs can be used to extract features. This enables the rapid and accurate construction of an autonomous, hierarchical system for extracting highly complex features of UAVs, achieving higher recognition accuracy, stronger anti-interference capabilities, and better generalization of UAVs.

[0061] (3): This invention, by comparing the overlapping flight paths of UAVs and identifying the positional relationship between UAVs and sensitive and no-fly zones, can meet the requirements for real-time position monitoring and dynamic control of different types of UAVs within the target controlled airspace. This allows for targeted signal intervention or forced control of the dynamic flight of local and external UAVs, preventing UAVs from continuing to fly within the target controlled airspace and causing uncontrollable harm. This reduces the negative impact of UAVs on the target controlled airspace and ensures the safety of the target controlled airspace. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the architecture of a full-band unmanned aerial vehicle (UAV) identification and control system based on deep learning, provided in Embodiment 1 of the present invention.

[0063] Figure 2 This is a flowchart illustrating a deep learning-based full-band drone identification and control method provided in Embodiment 2 of the present invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Example 1: Please refer to Figure 1 As shown in this embodiment, a deep learning-based full-band UAV identification and control system is applied to an air traffic control platform and includes:

[0066] The signal processing module uses the location coordinates of the signal receiving device as the airspace center to determine the target control airspace with airspace boundaries, and collects the raw analog signals of the entire frequency band in the target control airspace. After preprocessing the raw analog signals, it converts the raw analog signals into digital signals.

[0067] A signal receiving device refers to a multi-channel monitoring and receiving device for UAV signals that integrates one or more receiving antennas, enabling the signal receiving device to collect UAV signals. In this embodiment, the number of signal receiving devices is one, the number of receiving antennas on the signal receiving device is more than one, and the signal receiving device is deployed at a ground location where airspace control is required.

[0068] The target control airspace refers to the airspace area where drones need to be identified and controlled. The target control airspace also serves as the maximum spatial range for drone identification and control operations. Since the airspace is generally open and there are no actual physical boundaries, it is necessary to mark the airspace boundaries of the target control area in order to accurately carry out drone identification and control operations, thereby providing location constraints for subsequent drone identification and control operations.

[0069] When determining the target controlled airspace, it is necessary to first determine the airspace center of the target controlled area, and then expand outwards from the airspace center as a reference point to determine the target airspace.

[0070] Specifically, the steps for determining the target control area are as follows:

[0071] The real-time power of the signal receiving device is detected by a power sensor, and the power compliance rate is calculated by dividing the real-time power by the rated power.

[0072] The weather and terrain characteristics of the signal receiving equipment at the current moment are determined. These characteristics, along with the power compliance rate, are summarized into real-time factors. A factor lookup table is then used to determine the real-time receiving distance of the signal receiving equipment under these factors. Weather and terrain characteristics specifically represent the weather and terrain conditions of the signal receiving equipment's geographical location. Since different weather and terrain conditions can affect the performance of the signal receiving equipment to varying degrees, it is necessary to determine the current weather and terrain characteristics. Specifically, weather characteristics include sunny, rainy, and snowy days; terrain characteristics include plains, mountains, and hills.

[0073] The location coordinates of the signal receiving device are obtained by querying the BeiDou positioning system. The location coordinates are taken as the airspace center and the real-time receiving distance is taken as the radiation standard. Points that are one radiation standard away from the airspace center are recorded as boundary points, and F boundary points are obtained.

[0074] Connect any three boundary points that are not in the same direction to form a triangular grid. After summing up all the triangular grids, they are pieced together to form the airspace boundary, and the area within the airspace boundary is recorded as the target control area.

[0075] It should be noted that the airspace boundary is a physical boundary used to mark the spatial range covered by the target control area, which can provide a basis for judging whether the position of the UAV and the target control area are included or overlapped.

[0076] After obtaining the target control area, it is necessary to collect and analyze the electromagnetic wave signals of the drones entering the target control area. When determining whether the drone has entered the target control area, it is necessary to analyze and compare the drone's real-time position with the position of the airspace boundary.

[0077] In this embodiment, when determining whether a drone has entered the target control area, the drone's flight coordinates are queried in real time through the BeiDou positioning system. When the drone's flight coordinates coincide with the position coordinates of the airspace boundary, it can be determined that the drone has entered the target control area. At this time, it is necessary to collect and analyze the drone's electromagnetic wave signals.

[0078] The raw analog signal is the electromagnetic wave signal that exists in the target control area and can be collected by the signal receiving equipment. By collecting the raw analog signal of the whole frequency band, the electromagnetic wave signal of all frequency ranges in the target control airspace can be collected, thereby avoiding the phenomenon of missing electromagnetic wave signal collection and realizing the full range coverage collection effect of the UAV's electromagnetic wave signal.

[0079] Since the acquired raw analog signal is for the entire frequency band, it will also contain other electromagnetic signals that do not belong to the drone, including but not limited to Wi-Fi, Bluetooth, 4G / 5G and other electromagnetic signals. Therefore, it is necessary to preprocess the acquired raw analog signal and convert the preprocessed raw analog signal into a digital signal that can be directly used and analyzed.

[0080] Specifically, the digital signal conversion steps are as follows:

[0081] Based on the frequency magnitude, the electromagnetic wave frequencies in the communication field are divided into A communication frequency bands, and the operating frequency band of the receiving antenna on the signal receiving equipment is adjusted to fully cover the A communication frequency bands.

[0082] Electromagnetic wave signals in the target controlled airspace in A communication frequency bands are collected and denoted as the original analog signals;

[0083] The frequency range of the bandpass filter is set to be consistent with A communication frequency bands. The original analog signal is filtered by the bandpass filter, and the filtered original analog signal is linearly amplified by the low-noise amplifier to obtain a high-frequency radio frequency signal. If filtering is not performed, the strong out-of-band signal will overwhelm the subsequent amplifiers and digital-to-analog converters, making it impossible to extract the weak target signal. Because the noise figure of the low-noise amplifier itself has the greatest impact on the signal-to-noise ratio of the entire system, the low-noise amplifier is placed at the very beginning of the receiver chain.

[0084] The high-frequency radio frequency signal is input into the mixer to be down-frequency, and a low-frequency radio frequency signal is output.

[0085] The sampling clock update duration is used as the sampling duration of the analog-to-digital converter (ADC). Low-frequency radio frequency (RF) signals are sampled at intervals of one sampling duration, and all voltage values ​​are rounded to the nearest discrete level to output a digital signal. The level is represented by a binary digit, and the ADC output is a series of discrete digital samples representing the waveform of the original analog signal.

[0086] It should be noted that the purpose of the preprocessing operation on the original analog signal is to convert the high-frequency, low-power analog signal received by the antenna into a low-rate, digitized signal that can be processed by the digital signal processor. This allows the UAV's radio frequency signal to be converted from the analog domain to the digital domain, becoming a digital signal that can be directly processed by subsequent deep learning algorithms.

[0087] The time-frequency diagram construction module divides the digital signal into sub-signal sets labeled with sequence numbers, performs time-frequency analysis on the sub-signal sets, simulates time-frequency sub-graphs, and splices the time-frequency sub-graphs into a time-frequency diagram in sequence based on the sequence numbers.

[0088] After obtaining the digital signal, the digital signal can summarize and represent the electromagnetic wave signal in a continuous time period in the target control airspace. Since the time span corresponding to the digital signal is large, the time line corresponding to the digital signal is also long. In order to facilitate the accurate analysis of electromagnetic wave signals at different times and in shorter time periods, the digital signal needs to be divided according to the time line to generate a sub-signal set.

[0089] In this embodiment, a sub-signal set refers to a set of digital signals that are distinguished and summarized according to the order of their acquisition time. This allows the sub-signal set to effectively segment digital signals with a large time span and a large number of signals, and provides a direct basis for further processing of the digital signals.

[0090] When dividing a digital signal into sub-signal sets, it is necessary to label different sub-signal sets with corresponding timeline sequence numbers according to the time sequence of each digital signal, so that the sequence signals can serve as the basis for distinguishing and arranging different sub-signal sets.

[0091] Specifically, the steps for partitioning the sub-signal set are as follows:

[0092] The receiving time of each digital signal is marked by a timestamp, and all digital signals are arranged in chronological order according to the timeline.

[0093] The digital signal obtained by rounding is denoted as the first signal, and the digital signal obtained by rounding is denoted as the second signal. The number of the first signal and the number of the second signal are counted respectively, and denoted as the first value and the second value.

[0094] The signal difference is generated by taking the absolute value of the difference between the first and second values. The signal difference is then compared with the number of digital signals to calculate the unit spread ratio.

[0095] The formula for calculating the unit expansion ratio is:

[0096]

[0097] In the formula, KF bl For the unit expansion ratio, LZ d1 As the first value, LZ d2 The second value, XZ sl The number of digital signals;

[0098] The sampling duration is increased by one unit of amplitude to generate the set duration;

[0099] The formula for calculating the duration of the episode is:

[0100] SC jh=SC cy *(1+KF bl );

[0101] In the formula, SC jh For the duration of the collection, SC cy Sampling duration;

[0102] Starting from the first reception time and ending from the last reception time, and using the set duration as the set time period, B set time periods are divided between the start and end points; this can provide corresponding and accurate classification criteria for digital signals in different time periods.

[0103] The received digital signals within the set time period are divided sequentially to generate B sub-signal sets. Sequence numbers are then noted on the B sub-signal sets according to the chronological order of the timeline, with 1 as the first number.

[0104] In this embodiment, the sequence number is used to mark the temporal order of the sub-signal sets and to provide a basis for further processing of the subsequent sub-signal sets. Specifically, the sub-signal sets are numbered in ascending order, with 1 as the first number and B as the last number.

[0105] After dividing the digital signal into sub-signal sets, further time-frequency analysis can be performed on the digital signals in each sub-signal set. Based on the results of the time-frequency analysis, the energy density or intensity of the digital signal at different times and frequencies can be determined, thereby representing the power strength of each sub-signal set at different times and frequencies, and simulating the time-frequency sub-graph.

[0106] Specifically, the simulation method for the time-frequency subgraph is as follows:

[0107] According to the order of reception time, the digital signals in the sub-signal set are divided into C frame time periods, and the last moment of the previous frame time period is adjusted to overlap with the first moment of the next frame time period.

[0108] The digital signal within each of the C frame time periods is multiplied by a window function to window the digital signal. The windowed digital signal is then subjected to a Fast Fourier Transform (FFT) sequentially to output the amplitude and phase information of the digital signal. Windowing is used to suppress spectral leakage during the Fourier Transform, thereby obtaining clearer and more accurate spectral characteristics. Window functions include Hanning window, Hamming window, etc.

[0109] After summing the amplitude and phase information, a complex spectrum is generated. Then, the complex spectrum in the C frame time period is converted into C power spectra by performing modulo and squaring operations. Modulo is to calculate the modulus of each complex frequency point, and squaring is to calculate the square of each modulus value.

[0110] Using frequency index as rows, time frame index as columns, and power value as element value, arrange the power spectra of C frame time periods into a two-dimensional matrix, and then transform the two-dimensional matrix graph to obtain the time-frequency sub-graph;

[0111] By traversing B sub-signal sets in ascending order of sequence number, B time-frequency sub-graphs are simulated.

[0112] In this embodiment, each time-frequency sub-map can only represent the digital signals within the sub-signal set, and cannot represent all digital signals as a whole. Therefore, it is necessary to sequentially splice the time-frequency sub-maps of the B sub-signal sets to obtain a time-frequency map that can represent the digital signals in the target control airspace as a whole.

[0113] Specifically, when stitching together B time-frequency subgraphs into a time-frequency graph, firstly, according to the sequence number in ascending order, the time-frequency subgraphs corresponding to two adjacent sequence numbers are recorded as a subgraph unit, resulting in C-1 subgraph units; then, the portion of the time-frequency subgraph corresponding to the first moment of the next time-frequency subgraph in the subgraph unit is removed, and the previous time-frequency subgraph and the remaining next time-frequency subgraph are stitched together along the time axis to generate C-1 unit graphs; finally, the previous unit graph and the next unit graph are stitched together sequentially to obtain the time-frequency graph.

[0114] The deep learning module extracts multi-dimensional features from the time-frequency graph through a deep learning model. These multi-dimensional features include waveform features, modulation mode features, and spectrum features. The module also performs feature attribute analysis on these multi-dimensional features to identify the drone as either a local drone or an external drone.

[0115] After obtaining the time-frequency map, it can be used as a holistic representation of the electromagnetic wave signals emitted by UAVs within the target control airspace. At this time, the time-frequency map does not directly provide data for identifying UAV signals, but contains a large number of implicit detailed features of different dimensions. Therefore, it is necessary to identify and extract the multi-dimensional features in the time-frequency map.

[0116] Specifically, multi-dimensional features include waveform features, modulation features, and spectral features. Among them, waveform features are the most basic features, which deal with the instantaneous attributes of digital signals and are used to represent the characteristics of amplitude, phase change patterns, pulse shape, and period. Modulation features are an aggregation of waveform features and are the key to distinguishing different types of UAVs. Spectral features are the features observed after the signal is converted from the time domain to the frequency domain, including but not limited to fine spectral shape, spectral occupancy patterns, and multi-carrier features.

[0117] When extracting multi-dimensional features, it is necessary to combine deep learning in artificial intelligence technology to perform the extraction operation, so as to quickly and accurately extract the required waveform features, modulation mode features and spectrum features from the time-frequency graph containing a large number of complex and obscure features of different types.

[0118] Specifically, before extracting multi-dimensional features, the required deep learning model needs to be pre-trained so that the deep learning model can take the input time-frequency map as a basis, combine the preliminary feature extraction of shallow convolutional layers, and the preliminary feature fusion of middle and deep convolutional layers, and output the waveform features, modulation mode features and spectrum features corresponding to the time-frequency map.

[0119] It should be noted that deep learning models can be trained end-to-end to autonomously and hierarchically discover and build a highly complex feature extraction system for identifying drones directly from the most raw radio frequency data. This feature extraction system covers microscopic differences in multiple dimensions such as waveform, modulation, spectrum, and statistics, thereby achieving higher recognition accuracy, stronger anti-interference ability, and better generalization.

[0120] After extracting waveform features, modulation mode features, and spectrum features through deep learning models, it is necessary to analyze the feature attributes of waveform features, modulation mode features, and spectrum features to determine the specific type of UAV corresponding to the electromagnetic wave signal received in the target controlled airspace, and to determine whether the UAV entering the target controlled airspace is a local UAV or an external UAV based on the analysis results of the feature attributes.

[0121] Among them, the characteristic attributes are used to represent the specific type of UAV corresponding to waveform characteristics, modulation mode characteristics and spectrum characteristics, and serve as a direct basis for judging whether the UAV is a local UAV or an external UAV.

[0122] Local drones refer to drones that have been registered on the air traffic control platform and have a clear flight plan, while drones from outside the air traffic control platform refer to drones that have not been registered on the air traffic control platform and do not have a clear flight plan.

[0123] The steps for identifying local or out-of-territory drones are as follows:

[0124] The waveform features, modulation scheme features, and spectrum features are compared with the corresponding calibration features in the database for time-frequency analysis.

[0125] When the waveform characteristics match the calibrated waveform characteristics, it indicates that the UAV's fluctuation characteristics have been reported in the air traffic control platform, and the fluctuation characteristics are recorded as reported characteristics; the calibrated waveform characteristics refer to the waveform characteristics of UAVs that have been reported and pre-stored in the database.

[0126] When the modulation mode characteristics are consistent with the calibrated modulation mode characteristics, it means that the UAV's modulation mode characteristics have been reported in the air traffic control platform, and the modulation mode characteristics are recorded as reported characteristics; the calibrated modulation mode characteristics refer to the modulation mode characteristics of UAVs that have been reported and pre-stored in the database.

[0127] When the spectrum characteristics match the calibrated spectrum characteristics, it means that the UAV's spectrum characteristics have been reported in the air traffic control platform, and the spectrum characteristics are recorded as reported characteristics; the calibrated spectrum characteristics refer to the spectrum characteristics of UAVs that have been reported and stored in the database in advance.

[0128] The number of reported features of the drone is counted. When the number of reported features is 3, it means that the multi-dimensional features of the drone are all within the controllable range, and the drone is recorded as a local drone.

[0129] When the number of reported features is not 3, it indicates that the multi-dimensional features of the drone are not within the controllable range, and the drone is recorded as an off-domain drone.

[0130] It should be noted that when a drone is classified as an out-of-area drone, the number of reported characteristics may be any one of 0, 1, or 2. When such a drone enters the target controlled airspace, its flight trajectory and path may negatively affect the relevant equipment in the target controlled airspace. Therefore, it is necessary to differentiate between local drones and out-of-area drones for control.

[0131] The first instruction formulation module, when the drone is a local drone, collects the flight attitude data of the local drone, formulates the flight path of the local drone, compares the flight path with the planned path, and formulates the emergency control instructions for the local drone.

[0132] When the drone is a local drone, the probability of it causing negative impacts on relevant equipment within the target controlled airspace is relatively small. Therefore, it is only necessary to monitor the flight status of the local drone in real time.

[0133] In order to achieve real-time monitoring of the flight status of local drones, it is necessary to collect the flight attitude data of local drones and use the flight attitude data to comprehensively represent a series of data such as the flight speed, heading, and trajectory of local drones in the target controlled airspace.

[0134] Specifically, flight attitude data includes flight speed, flight heading, and flight coordinates; when collecting flight attitude data, it is obtained through real-time querying via the BeiDou positioning system.

[0135] After obtaining the drone's flight attitude data, a flight path that can intuitively represent the historical flight trajectory of the local drone can be constructed based on the flight speed, flight heading, and flight coordinates.

[0136] Specifically, the steps for determining the flight path are as follows:

[0137] The moment when the local drone first enters the target controlled airspace is recorded as the start time, and the time period from the start time to the current time is recorded as the flight period.

[0138] The flight coordinates of the local drone at D times during the flight period are retrieved one by one using the BeiDou positioning system. The points where the D flight coordinates are located are marked one by one on the satellite map. The D points are then connected in sequence to generate the initial path.

[0139] The flight speed and heading of the local drone at D time points are retrieved one by one, and the flight speed and heading at the same time point are bound together to generate D flight parameters;

[0140] The D flight parameters are noted on the D points respectively, causing the initial path to be converted into a flight path.

[0141] After obtaining the flight path of the local drone, it is necessary to compare the flight path of the local drone with the pre-planned path to determine whether the flight path of the local drone overlaps with or deviates from the pre-planned path. Based on the result of the overlap comparison, emergency control instructions for the local drone are formulated.

[0142] Specifically, emergency control instructions include instructions to continue flight and instructions to adjust flight.

[0143] The steps for formulating emergency control instructions are as follows:

[0144] The planned path of the local drone is retrieved through the air traffic control platform. The part of the planned path corresponding to the flight path is recorded as the target path, and E trajectory points on the target path are marked.

[0145] Find the position coordinates of the points on the flight path that correspond to the E trajectory points one by one, record the trajectory points that have the same position coordinates as the points as coincident points, and count the number of coincident points.

[0146] Divide the number of overlapping points by the number of trajectory points to calculate the overlap ratio, and analyze the magnitude of the overlap ratio.

[0147] When the overlap ratio is greater than or equal to the preset overlap threshold, it indicates that there are many trajectory points that overlap with the flight path and the planned path, and the degree of overlap of the flight path is high. In this case, a continuous flight command is issued. The preset overlap threshold is the critical value of the overlap ratio corresponding to the issuance of a continuous flight command or flight adjustment command. It serves as a numerical basis for judging the degree of overlap between the flight path and the target path. In this embodiment, the preset overlap threshold can be set according to actual needs. For example, the preset overlap threshold is 95%.

[0148] When the overlap ratio is less than the preset overlap threshold, it indicates that the number of trajectory points that overlap with the flight path and the planned path is small, and the degree of overlap of the flight path is low. In this case, a flight adjustment command is issued.

[0149] It should be noted that after a continuous flight instruction is issued, the air traffic control platform will not intervene or control the local drone, and the local drone can continue to fly according to its current flight status. However, when a flight adjustment instruction is issued, the air traffic control platform needs to intervene and control the local drone. The air traffic control platform sends relevant instructions to the local drone to adjust and control its flight path.

[0150] The second instruction formulation module, when the UAV is an out-of-domain UAV, divides the target controlled airspace into sensitive airspace and no-fly airspace, judges the positional relationship between the out-of-domain UAV and the sensitive airspace and no-fly airspace in real time, and formulates sensitive driving instructions or electromagnetic shielding instructions for the out-of-domain UAV.

[0151] When a drone is an out-of-area drone, the probability of it negatively impacting relevant equipment within the target controlled airspace is relatively high. Therefore, it is necessary to monitor and dynamically control the flight status of local drones in real time.

[0152] Because the different locations of drones from outside the target airspace can have varying degrees of impact on related equipment, air traffic control platforms need to implement different types of control measures for drones from outside the target airspace. Therefore, it is necessary to analyze the specific circumstances of drones from outside the target airspace, further divide the target airspace, and divide the target airspace into sensitive airspace and no-fly airspace.

[0153] Specifically, sensitive airspace refers to the space area within the target controlled airspace that is relatively far from the center of the airspace, while no-fly airspace refers to the space area within the target controlled airspace that is relatively close to the center of the airspace.

[0154] In this embodiment, when dividing the target controlled airspace into sensitive airspace and no-fly airspace, two-thirds of the real-time receiving distance is used as the division distance, and the portion of the target controlled airspace that is one division distance from the airspace center is recorded as no-fly airspace, while the remaining portion of the target controlled airspace is recorded as sensitive airspace.

[0155] After dividing the target control airspace into sensitive airspace and no-fly airspace, the real-time position of UAVs outside the airspace can be collected, and the relative positional relationship between the real-time position and the sensitive airspace and no-fly airspace can be determined. Based on the different positional relationships, corresponding control instructions can be formulated.

[0156] Positional relationship is used to indicate whether the current position of an out-of-domain UAV is within a sensitive airspace or a no-fly zone. In this embodiment, the positional relationship between the out-of-domain UAV and the sensitive airspace and no-fly zone includes being within the sensitive airspace and being within the no-fly zone.

[0157] Specifically, the control instructions include sensitive expulsion instructions and electromagnetic shielding instructions. Sensitive expulsion instructions are issued by the air traffic control platform to expel drones from outside the region when they enter sensitive airspace, while electromagnetic shielding instructions are issued by the air traffic control platform to shield drones from outside the region when they enter no-fly airspace.

[0158] Specifically, the steps for formulating sensitive expulsion instructions or electromagnetic shielding instructions are as follows:

[0159] The location coordinates of all points in sensitive areas and no-fly zones were queried one by one using the BeiDou positioning system. The location coordinates of sensitive areas and no-fly zones were then summarized to generate sensitive coordinate databases and no-fly coordinate databases.

[0160] The system retrieves the flight coordinates of the out-of-domain drone at the current moment and compares these coordinates with the position coordinates in the sensitive coordinate database and the no-fly zone coordinate database.

[0161] When the flight coordinates match the position coordinates in the sensitive coordinate database, the UAV outside the domain is located inside the sensitive airspace at the current moment, and a sensitive driving-away command is issued.

[0162] When the flight coordinates match the position coordinates in the no-fly zone coordinate database, and the drone from outside the zone is located within the no-fly zone at the current moment, an electromagnetic shielding command is issued.

[0163] In this embodiment, after a sensitive expulsion command is issued, if the air traffic control platform needs to perform mild interference and control measures on the out-of-area drone, it sends an expulsion message to the out-of-area drone to prevent it from continuing to fly in the sensitive airspace and to drive it away from the target controlled airspace as soon as possible. After an electromagnetic shielding command is issued, if the air traffic control platform needs to perform severe interference and control measures on the out-of-area drone, it sends an electromagnetic shielding message to the out-of-area drone to shield the transmission channel between the out-of-area drone and external electromagnetic wave signals, cut off the control link of the out-of-area drone, and cause the out-of-area drone to lose control or even crash at the first time, thereby preventing more serious negative impacts.

[0164] Example 2: Please refer to Figure 2 As shown, parts not described in detail in this embodiment are described in Embodiment 1. This embodiment provides a deep learning-based full-band UAV identification and control method, applied to an air traffic control platform, and implemented based on a deep learning-based full-band UAV identification and control system, including:

[0165] S01: Using the signal receiving device as the airspace center, determine the target control airspace, collect the raw analog signals of the entire frequency band in the target control airspace, and process and convert the raw analog signals into digital signals;

[0166] S02: Divide the digital signal into sub-signal sets labeled with sequence numbers, perform time-frequency analysis on the sub-signal sets, simulate the time-frequency sub-graphs corresponding to the sub-signal sets, and stitch the time-frequency sub-graphs into a time-frequency graph in sequence based on the sequence numbers;

[0167] S03: Extract multi-dimensional features from the time-frequency graph using a deep learning model, and perform feature attribute analysis on the multi-dimensional features to identify the drone as a local drone or an external drone; if it is a local drone, proceed to S04; if it is an external drone, proceed to S05.

[0168] S04: Collect the flight speed, flight heading and flight coordinates of the local UAV, formulate the flight path of the local UAV, compare the flight path with the planned path, and formulate continuous flight command or flight adjustment command.

[0169] S05: Divide the target control airspace into sensitive airspace and no-fly airspace, determine the positional relationship between the UAV outside the airspace and the sensitive airspace and no-fly airspace in real time, and formulate sensitive driving-away instructions or electromagnetic shielding instructions.

[0170] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based full-band unmanned aerial vehicle (UAV) identification and control system, applied to an air traffic control platform, characterized in that, include: The signal processing module, with the signal receiving device as the airspace center, determines the target control airspace, collects the raw analog signals in the full frequency band of the target control airspace, and processes and converts the raw analog signals into digital signals. The time-frequency diagram construction module divides the digital signal into sub-signal sets labeled with sequence numbers, performs time-frequency analysis on the sub-signal sets, simulates the time-frequency sub-graphs corresponding to the sub-signal sets, and splices the time-frequency sub-graphs into a time-frequency diagram in sequence based on the sequence numbers. The steps for partitioning the sub-signal set are as follows: The digital signal obtained by rounding is denoted as the first signal, and the digital signal obtained by rounding is denoted as the second signal. The number of the first signal and the number of the second signal are counted respectively, and denoted as the first value and the second value. The signal difference is generated by taking the absolute value of the difference between the first and second values. The signal difference is then compared with the number of digital signals to calculate the unit spread ratio. The sampling duration is increased by one unit of amplitude to generate the set duration; Mark the reception times of all digital signals, and divide the time interval between the first and last reception times into B sets; The digital signals received within the set time period are sequentially aggregated to generate B sub-signal sets, and each sub-signal set is labeled with a sequence number starting with 1. The simulation method for time-frequency subgraphs is as follows: According to the order of reception time, the digital signal of the sub-signal set is divided into C frame time periods, and the last moment of the previous frame time period is adjusted to overlap with the first moment of the next frame time period. The digital signals for each of the C frame time periods are multiplied by a window function to form a window, and the windowed digital signals are then subjected to a Fast Fourier Transform sequentially to output amplitude and phase information. After summing the amplitude and phase information, a complex spectrum is generated, and modulo and squaring operations are performed on the complex spectrum of C frame time periods to convert the C complex spectrum into C power spectra. Using frequency index as the row, time frame index as the column, and power value as the element value, arrange the power spectrum of C frame time periods into a two-dimensional matrix, and transform the two-dimensional matrix graph to obtain the time-frequency sub-graph. After traversing B sub-signal sets in ascending order of sequence number, simulate B time-frequency sub-graphs. The deep learning module extracts multi-dimensional features from the time-frequency graph through a deep learning model. These multi-dimensional features include waveform features, modulation mode features, and spectrum features. The module also performs feature attribute analysis on these multi-dimensional features to identify the drone as either a local drone or an external drone. The first command formulation module collects the flight speed, flight heading and flight coordinates of the local UAV, formulates the flight path of the local UAV, compares the flight path with the planned path, and formulates the continuous flight command or flight adjustment command. The second instruction formulation module divides the target control airspace into sensitive airspace and no-fly airspace, determines the positional relationship between the UAV outside the airspace and the sensitive airspace and no-fly airspace in real time, and formulates sensitive driving-away instructions or electromagnetic shielding instructions.

2. The deep learning-based full-band UAV identification and control system according to claim 1, characterized in that, The steps for determining the target control area are as follows: The real-time power of the signal receiving device is detected by a power sensor, and the power compliance rate is calculated by dividing the real-time power by the rated power. Determine the weather and terrain characteristics of the signal receiving equipment at the current moment, summarize the weather characteristics, terrain characteristics, and power compliance rate into real-time factors, and look up the real-time receiving distance of the signal receiving equipment under the real-time factors through the factor comparison table; The location coordinates of the signal receiving device are obtained by querying the BeiDou positioning system. The location coordinates are taken as the airspace center and the real-time receiving distance is taken as the radiation standard. Points that are one radiation standard away from the airspace center are recorded as boundary points, and F boundary points are obtained. Connect any three boundary points that are not in the same direction to form a triangular grid. After summing up all the triangular grids, they are pieced together to form the airspace boundary, and the area within the airspace boundary is recorded as the target control area.

3. The deep learning-based full-band UAV identification and control system according to claim 2, characterized in that, The steps for digital signal conversion are as follows: Adjust the operating frequency band of the receiving antenna on the signal receiving equipment to fully cover the A communication frequency bands in the communication field, collect the electromagnetic wave signals of the target control airspace in the A communication frequency bands, and record them as the original analog signals; The frequency range of the bandpass filter is set to be consistent with A communication frequency bands. The original analog signal is filtered by the bandpass filter, and the filtered original analog signal is linearly amplified by a low-noise amplifier to obtain a high-frequency radio frequency signal. The high-frequency radio frequency signal is input into the mixer to be down-frequency, and a low-frequency radio frequency signal is output. The sampling clock update duration is used as the sampling duration of the analog-to-digital converter. The voltage values ​​of the low-frequency radio frequency signal are sampled at intervals of one sampling duration, and all voltage values ​​are rounded to the nearest discrete level to output a digital signal.

4. The deep learning-based full-band UAV identification and control system according to claim 3, characterized in that, The steps for identifying local or out-of-territory drones are as follows: When the waveform characteristics are consistent with the calibrated waveform characteristics, the fluctuation characteristics are recorded as reported characteristics; When the modulation mode characteristics match the calibrated modulation mode characteristics, the modulation mode characteristics are recorded as reported characteristics; When the spectral characteristics match the calibrated spectral characteristics, the spectral characteristics are recorded as reported characteristics; The number of reported features of the drone is counted. When the number of reported features is 3, the drone is recorded as a local drone. When the number of reported features is not 3, the drone will be recorded as an off-domain drone.

5. A deep learning-based full-band UAV identification and control system according to claim 4, characterized in that, The steps for determining the flight path are as follows: The moment when the local drone first enters the target controlled airspace is recorded as the start time, and the time period from the start time to the current time is recorded as the flight period. The flight coordinates of the local drone at D times during the flight period are retrieved one by one using the BeiDou positioning system. The points where the D flight coordinates are located are marked one by one on the satellite map. The D points are then connected in sequence to generate the initial path. The flight speed and heading of the local drone at D times are queried one by one. After binding the flight speed and heading at the same time, D flight parameters are generated. The D flight parameters are noted on D points respectively, so that the initial path is converted into a flight path.

6. A deep learning-based full-band UAV identification and control system according to claim 5, characterized in that, The steps for formulating a continuous flight instruction or a flight adjustment instruction are as follows: The planned path of the local drone is retrieved through the air traffic control platform. The part of the planned path corresponding to the flight path is recorded as the target path, and E trajectory points on the target path are marked. Find the position coordinates of the points on the flight path that correspond to the E trajectory points one by one, record the trajectory points that have the same position coordinates as the points as coincidence points, count the number of coincidence points, divide the number of coincidence points by the number of trajectory points, and calculate the overlap ratio. When the overlap ratio is greater than or equal to the preset overlap threshold, a continuous flight command is issued. When the overlap ratio is less than the preset overlap threshold, a flight adjustment command is generated.

7. A deep learning-based full-band unmanned aerial vehicle (UAV) identification and control system according to claim 6, characterized in that, The steps for formulating sensitive expulsion commands or electromagnetic shielding commands are as follows: The location coordinates of all points in sensitive areas and no-fly zones were queried one by one using the BeiDou positioning system. The location coordinates of sensitive areas and no-fly zones were then summarized to generate sensitive coordinate databases and no-fly coordinate databases. The system retrieves the flight coordinates of an out-of-domain drone at the current moment. If the flight coordinates match the location coordinates in the sensitive coordinate database, the out-of-domain drone is located within the sensitive airspace at the current moment, and a sensitive expulsion command is issued. When the flight coordinates match the position coordinates in the no-fly zone coordinate database, the UAV outside the zone is located inside the no-fly zone at the current moment, and an electromagnetic shielding command is issued.

8. A deep learning-based full-band UAV identification and control method, applied to an air traffic control platform, implemented based on any one of claims 1-7, characterized in that... include: S01: Using the signal receiving device as the airspace center, determine the target control airspace, collect the raw analog signals of the entire frequency band in the target control airspace, and process and convert the raw analog signals into digital signals; S02: Divide the digital signal into sub-signal sets labeled with sequence numbers, perform time-frequency analysis on the sub-signal sets, simulate the time-frequency sub-graphs corresponding to the sub-signal sets, and stitch the time-frequency sub-graphs into a time-frequency graph in sequence based on the sequence numbers; S03: Extract multi-dimensional features from the time-frequency graph using a deep learning model, and perform feature attribute analysis on the multi-dimensional features to identify the drone as a local drone or an external drone; if it is a local drone, proceed to S04; if it is an external drone, proceed to S05. S04: Collect the flight speed, flight heading and flight coordinates of the local UAV, formulate the flight path of the local UAV, compare the flight path with the planned path, and formulate continuous flight command or flight adjustment command. S05: Divide the target control airspace into sensitive airspace and no-fly airspace, determine the positional relationship between the UAV outside the airspace and the sensitive airspace and no-fly airspace in real time, and formulate sensitive driving-away instructions or electromagnetic shielding instructions.

Images