A UAV remote detection method based on network transmission

By dividing the low-altitude UAV detection into gridded control units and constructing a transmission mapping topology, and adaptively selecting feature levels and coding rates, multimodal fusion processing and active detection are achieved. This solves the problems of high transmission costs and high false alarm rates in existing technologies, and improves detection sensitivity and accuracy.

CN123123326APending Publication Date: 2026-11-17BEIJING RONG MEDIA ZHIZHONG NETWORK TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611278551.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing low-altitude drone detection technologies suffer from several problems in wide-area low-altitude control scenarios, including limited edge detection node capabilities, insufficient multimodal deep fusion, high transmission costs, high false alarm and false detection rates, and inability to actively detect silent targets.

Method used

The airspace is divided into standardized grid control units, and integrated radio frequency, acoustic, and photoelectric sensing nodes are deployed to construct a transmission mapping topology. Feature levels and coding rates are adaptively selected, and multimodal fusion processing and feature verification are performed in the cloud to actively detect suspected targets and dynamically adjust the detection model parameters.

Benefits of technology

It reduced transmission costs, improved detection sensitivity and accuracy, reduced false alarm rate, and ensured real-time performance and rapid response to high-risk targets.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123326A_ABST
    Figure CN123123326A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of low-altitude unmanned aerial vehicle (UAV) control technology and discloses a remote UAV detection method based on network transmission. According to low-altitude air traffic control rules, the airspace is divided into three levels: no-fly zone, restricted zone, and general monitoring zone. Each level of airspace is further subdivided into standardized gridded control units. Within each unit, detection nodes integrating radio frequency, acoustic, and photoelectric sensing are deployed. Real-time data collection of link quality parameters from each detection node to the cloud, such as end-to-end latency, available bandwidth, packet loss rate, and latency jitter, is performed. A transmission mapping topology is constructed based on the control level and real-time situation. Dedicated transmission slices are allocated to higher-level units, while shared link scheduling is used for general units. The detection nodes extract three types of features from multimodal signals and bind them to control service weights. The transmission feature level and coding rate are adaptively selected based on network status and control level. This reduces the bandwidth overhead and transmission cost of full data return.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of low-altitude unmanned aerial vehicle (UAV) control technology, specifically a remote detection method for UAVs based on network transmission. Background Technology

[0002] Low-altitude drone detection and control is a crucial technological area in airspace management. Existing solutions often employ an architecture of edge detection nodes and network-based result transmission, using methods such as RFID, electro-optical radar, and acoustic detection to sense drones and then relaying the results back to the management platform. However, in practical applications of wide-area low-altitude control, existing technologies still face the following challenges: Edge detection nodes are limited by power consumption, size and deployment cost, and can only run lightweight single-modal detection algorithms. They cannot achieve deep multimodal fusion and weak signal accumulation enhancement, and their detection rate for small, slow, and unauthorized drones is insufficient. If the raw sensing data is transmitted back to the cloud for processing, it will result in bandwidth overhead of hundreds of megabits. When deployed in a wide area, the transmission cost and transmission latency cannot meet the real-time requirements of the control scenario.

[0003] Current solutions only transmit the final target position and velocity data, while the original signal features and intermediate semantic features are directly discarded at the edge. The cloud cannot perform cross-node feature-level association and fusion. Faced with complex scenarios such as multipath reflections from urban buildings, interference from civilian co-channel signals, and dense flight of multiple drones, it is impossible to eliminate false targets through multi-view feature verification or improve the detection sensitivity of weak signal targets through cross-node signal accumulation. This results in persistently high false alarm and missed detection rates in complex low-altitude environments.

[0004] The existing network transmission serves only as a passive channel for data backhaul, completely decoupled from the detection process, and cannot dynamically schedule transmission resources according to detection needs. When a suspected silent target is detected, the transmission bandwidth and latency of the corresponding node cannot be prioritized; for silent flying drones without radio frequency signals, it is also impossible to trigger active detection by multiple nodes through network coordination, resulting in a natural detection blind spot in the passive detection mode. Summary of the Invention

[0005] The purpose of this invention is to provide a remote detection method for unmanned aerial vehicles (UAVs) based on network transmission, so as to solve one or more problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a remote detection method for unmanned aerial vehicles (UAVs) based on network transmission, comprising the following specific steps: Furthermore, according to the low-altitude air traffic control rules, the target control airspace is divided into three levels of control airspace: no-fly zone, restricted zone, and general surveillance zone. The control airspace is then further subdivided into standardized grid control units, and each grid control unit is equipped with integrated radio frequency sensing, acoustic sensing, and photoelectric sensing detection nodes. The network link quality parameters from the detection node to the cloud control and processing platform are collected in real time, including end-to-end latency, available bandwidth, packet loss rate and latency jitter. Combined with the control level and real-time control status of the gridded control unit, the transmission mapping topology between the detection node and the cloud control and processing platform is constructed. For high-level grid-based control units corresponding to no-fly zones and restricted control zones, low-latency, high-reliability dedicated control transmission slices are allocated; for grid-based control units corresponding to general surveillance zones, shared link scheduling is adopted.

[0007] Furthermore, the detection node extracts features from the collected radio frequency, acoustic, and photoelectric signals, and sequentially generates feature data at three levels: original signal compression features, primary target candidate features, and deep semantic features. Each layer of features is pre-bound with corresponding regulatory business weights. The original signal compression feature corresponds to the regulatory-level signal tracing requirements, the primary target candidate feature corresponds to the regulatory-level early warning screening requirements, and the deep semantic feature corresponds to the regulatory-level target identity verification requirements. Based on the current real-time network bandwidth conditions of the detection node and the control level and real-time control situation weight of its grid-based control unit, the feature level and encoding rate to be transmitted are selected. When network link bandwidth is sufficient, latency and packet loss rate meet requirements, and the control level is high, deep semantic features are transmitted; when network link bandwidth is limited, latency jitter is large, and the control level is low, high compression ratio of raw signal compression features are transmitted, and the cloud-based control processing platform completes the deep feature extraction and target recognition calculation.

[0008] Furthermore, the cloud-based control and processing platform receives feature data transmitted by detection nodes within the same gridded control unit. Based on the timestamp of the unified air traffic control time reference, it first performs joint calculation of spatiotemporal parameters such as arrival time difference, arrival angle difference, and Doppler frequency shift on the feature data of different detection nodes to complete the spatiotemporal alignment of the detection node features, and then carries out multimodal fusion processing of feature dimensions. During the multimodal fusion process, verification is carried out simultaneously. Using airspace control flight rules as the constraint dimension, a compliance verification logic for target flight behavior is constructed. Combining whether the target trajectory breaks through the controlled airspace boundary and whether the flight speed and altitude comply with the flight rules of the corresponding controlled airspace, a target association matrix is ​​constructed together through spatiotemporal consistency constraints and airspace control flight rule constraints. False targets generated by multipath reflection and co-frequency interference are eliminated, and the trajectory and signal feature labels of suspected UAV targets that conform to the regulatory data format are output.

[0009] Furthermore, the cloud-based control and processing platform makes judgments based on the target trajectory and multi-dimensional feature tags output by the multi-modal fusion. When the acoustic and photoelectric sensing modes both detect target signals that match the flight characteristics of the UAV, and the radio frequency sensing mode does not capture the corresponding UAV flight control and image transmission radio frequency signals, and the target is within the controlled airspace warning range, it is determined to be a silent suspected target. When it is determined that the target has a tendency to break through the control boundary, an active detection control command is generated. Through the network's time slot scheduling mechanism, the transmission time slot resources of the control sector to which the corresponding grid control unit belongs are reconstructed. Some idle time slots originally used for issuing control commands are reused as dedicated transmission time slots for active detection commands and echo data. The active detection control commands are preferentially issued to the detection nodes within the corresponding grid control unit. The detection nodes are controlled to emit directional radio frequency excitation signals and acoustic detection pulses that comply with airspace electromagnetic control requirements. At the same time, dedicated transmission time slots for control are reserved for adjacent detection nodes for real-time transmission of active detection echo data. After the echo data is transmitted back to the cloud-based control processing platform, it is incorporated into the feature fusion and parameter calculation process of the corresponding target to improve the feature information of the silent target.

[0010] Furthermore, the cloud-based control and processing platform will match and verify the calculated target trajectory and signal characteristics with the UAV control identity feature database to identify the target's device model, signal pattern, and control terminal feature identifier, and simultaneously match the target's compliance registration status and flight permissions. Simultaneously, the same target captured by the detection node is subject to dual verification of feature consistency and control behavior consistency, eliminating false targets with mismatched features or flight behavior that does not conform to control logic; Output target detection results with regulatory attribute tags, which include four categories: compliant flight, unregistered flight, suspected illegal flight, and key areas of regulatory concern.

[0011] Furthermore, the cloud-based control and processing platform generates detection model parameters based on UAV target samples, false detection samples, and interference samples collected in specific control scenarios. The detection models are classified according to control scenarios, including multipath interference adaptation models for airport control areas, dense target adaptation models for urban control areas, and silent target adaptation models for border control areas. Through the network's breakpoint resume and fragmented transmission mechanism, the detection model parameters are fragmented and distributed to the detection nodes within the corresponding controlled airspace in descending order of airspace control level. Without interrupting normal control and detection operations, the detection node uses a parameter hot-loading method to complete the online incremental update of the local detection model.

[0012] Furthermore, the cloud-based control and processing platform performs three-level transmission scheduling of detection results—control handling level, control monitoring level, and control archiving level—based on the control attribute tags and corresponding control levels of the targets. The targets at the control handling level are directly transmitted to the control command terminal through a low-latency dedicated control link, and the corresponding target situation information is simultaneously pushed to the surrounding detection nodes to improve the control monitoring level of the corresponding area. The targets at the control monitoring level are pushed to the control terminal through the control monitoring link, and the targets at the control archiving level are returned and archived through a regular link. By receiving detection effect feedback data from detection nodes and control response feedback data from the control command terminal via the network, and using control response success rate, number of control malfunctions caused by false alarms, and control response delay for silent targets as core optimization indicators, the gridded control unit division, detection node topology, transmission resource configuration, and detection model weights of the corresponding controlled airspace are adjusted.

[0013] The beneficial effects of this invention are as follows: 1. This invention divides the airspace into three levels of control based on low-altitude air traffic control rules: no-fly zone, restricted zone, and general surveillance zone. Each level of airspace is further subdivided into standardized gridded control units. Within each unit, integrated radio frequency, acoustic, and photoelectric sensing detection nodes are deployed. Real-time data collection of link quality parameters from each detection node to the cloud, including end-to-end latency, available bandwidth, packet loss rate, and latency jitter, is used. A transmission mapping topology is constructed based on control level and real-time situation, allocating dedicated transmission slices to high-level units and using shared link scheduling for general units. Detection nodes extract three types of features from multimodal signals and bind them to control service weights. The transmission feature level and coding rate are adaptively selected based on network status and control level. This reduces the bandwidth overhead and transmission cost of full data return, balancing transmission reliability in high-control-level areas with resource utilization in ordinary areas, and maintaining stable detection data output under different network conditions.

[0014] 2. The cloud-based air traffic control processing platform of this invention receives feature data uploaded by multiple detection nodes within the same gridded air traffic control unit. First, based on the timestamp of the unified air traffic control time reference, it jointly calculates the spatiotemporal parameters of arrival time difference, arrival angle difference, and Doppler frequency shift for data from different nodes, completing the spatiotemporal alignment of multi-node features. Then, it performs multimodal fusion processing of feature dimensions, combining airspace control flight rules, and constructs a target association matrix with spatiotemporal consistency and flight compliance as constraints, eliminating false targets caused by multipath reflection and co-frequency interference. When a silent suspected target is identified, it reconstructs transmission time slot resources through time slot scheduling, reuses idle time slots to issue commands, controls nodes to transmit excitation signals, and reserves time slots to transmit echoes, supplementing the fusion calculation process. This reduces the false alarm rate in complex environments, improves target positioning accuracy and weak signal target detection sensitivity, and ensures the real-time performance of active detection services.

[0015] 3. The cloud-based control and processing platform of this invention uses UAV target samples, false detection samples, and interference samples collected in dedicated control scenarios such as airports, cities, and borders to train detection model parameters adapted to multiple scenarios. Through network breakpoint resume and fragmented transmission mechanisms, the model parameters are fragmented and distributed to the detection nodes in the corresponding coverage areas in descending order of airspace control level. The local model is updated online incrementally without interrupting normal business by hot loading of parameters. At the same time, a three-level transmission scheduling is performed according to the target control level, receiving detection and handling feedback data, and dynamically adjusting the airspace grid, node topology, transmission resources, and model weights based on core indicators. It can quickly adapt to the environmental interference characteristics of different areas, continuously improve detection accuracy, and shorten the control response latency for high-risk targets. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall detection and control closed-loop process of this invention. Figure 2 This is a flowchart of the silent target active detection execution sub-process of the present invention. Detailed Implementation

[0017] 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.

[0018] like Figures 1 to 2 As shown, this embodiment of the invention provides a remote detection method for unmanned aerial vehicles (UAVs) based on network transmission. This embodiment takes the airspace control area of ​​a 4E-level civil transport airport as a fixed application scenario and implements the entire process around the detection and control of low-altitude UAVs in the airport's airspace.

[0019] According to the Civil Aviation Airport Airspace Protection Management Rules, this embodiment divides the airport airspace control area into three levels: the runway end takeoff and landing zone and the area within 1 kilometer on both sides of the runway centerline are no-fly zones; the airspace corresponding to the airport's takeoff and landing routes is a restricted zone; and the airspace extension area within 5 kilometers of the airport's perimeter is a general surveillance zone. The airspace control areas at each level are further subdivided into standardized grid control units at equal intervals of latitude and longitude. Each grid control unit is equipped with no less than 3 integrated detection nodes in a distributed triangular layout that integrate radio frequency sensing, acoustic sensing, and photoelectric sensing.

[0020] In this embodiment, the side length of a single grid in the no-fly zone is set to 500 meters, the side length of a single grid in the restricted zone is set to 1000 meters, and the side length of a single grid in the general monitoring zone is set to 2000 meters. The detection nodes in each grid control unit are arranged in a distributed triangular layout and are deployed at fixed high points such as airport perimeter monitoring towers, navigation light towers, or terminal roof support columns at different locations within the grid. The radio frequency sensing radius of a single node is not less than 0.7 times the side length of the corresponding grid, and the sensing coverage of a single node forms an overlap area of ​​not less than 20% with the boundary of the adjacent grid. This is used to offset the multipath shading effect caused by the glass curtain wall of the terminal building and the metal roof of the apron, and to ensure the effective observation baseline for multi-node joint positioning.

[0021] Real-time network link quality parameters from each detection node to the airport cloud control processing platform are collected, including end-to-end latency, available bandwidth, packet loss rate, and latency jitter. Combined with the control level of the gridded control unit and the real-time control status of flight takeoffs and landings, a transmission mapping topology between the detection nodes and the cloud control processing platform is constructed.

[0022] In this embodiment, the transmission bearer network adopts a dedicated 5G private network for the airport + fiber optic backup links. The transmission mapping topology uses the cloud-based control processing platform deployed in the airport information center as the core node and the detection nodes of each grid unit as edge nodes. QoS priority identifiers are assigned to each link according to link quality parameters and control levels. The links corresponding to the no-fly zone are set to the highest priority and mapped to the uRLLC dedicated logical slice of the 5G private network. A hard guarantee of 10Mbps dedicated bandwidth and time slots is achieved through the RSVP-TE resource reservation protocol. The links corresponding to the general surveillance zone adopt shared link scheduling. Based on the DBA dynamic bandwidth allocation mechanism, a shared bandwidth pool of 50Mbps is dynamically allocated among multiple nodes.

[0023] In this embodiment, the transmission mapping topology is set with two modes: fixed-period update and event-triggered update. Fixed-period update refreshes the mapping relationship of all links every 30 seconds. Event-triggered update is triggered immediately when the packet loss rate of the probe node link increases by more than 5% for three consecutive sampling periods (sampling period is 1 second), the flight approach / takeoff control level is upgraded, or a suspected black flight target appears. It only updates the link mapping and resource allocation of the affected sectors.

[0024] For high-level grid control units corresponding to runway-end no-fly zones and takeoff and landing restricted zones, low-latency, high-reliability dedicated control transmission slices are allocated; for grid control units corresponding to general surveillance zones, shared link scheduling is adopted.

[0025] The detection node extracts features from the collected radio frequency, acoustic, and photoelectric signals, and generates feature data at three levels in sequence: original signal compression features, primary target candidate features, and deep semantic features.

[0026] In this embodiment, feature extraction is implemented using a three-level nested lightweight network adapted to embedded edge computing power: the first layer is a signal compression coding layer, which uses 4-layer 1D convolution, 4-layer 1D convolution, and lightweight MobileNetV2 backbone 2D convolution to downsample and compress and encode RF IQ sampling signals in the 30MHz~6GHz band, acoustic time-domain sampling signals in the 20Hz~20kHz band, and 1080P photoelectric image frames, respectively. The input is the original perceptual sampling data, and the output is the original signal compression features after dimensional compression. The middle layer is the candidate target detection layer. Based on compressed features, it extracts the time domain, frequency domain, and spatial domain candidate boxes of potential targets through sliding window detection and adaptive threshold screening. The input is the original signal compressed features, and the output is the primary target candidate features containing target orientation, signal energy, and initial screening confidence. The deep layer is a semantic recognition layer. It uses an 8-layer residual module in conjunction with a channel attention module to extract high-dimensional semantic information such as the signal modulation paradigm, spectral texture contour, and photoelectric morphological features of the target. The input is the local feature slices corresponding to the primary target candidate features, and the output is the deep semantic features.

[0027] In this embodiment, the dimension of the original signal compression feature is set to 256, the confidence threshold of the primary target candidate feature is set to 0.6, and the output dimension of the deep semantic feature is set to 128. Both the threshold and the dimension are adapted to the suppression requirements of background interference such as DME rangefinders and VHF voice communication in airport environments.

[0028] Each layer of features is pre-bound with corresponding regulatory business weights: the original signal compression feature corresponds to the regulatory-level signal tracing and interference analysis requirements, the primary target candidate feature corresponds to the regulatory-level early warning screening requirements, and the deep semantic feature corresponds to the regulatory-level target identity verification requirements.

[0029] In this embodiment, each feature level is set with a basic weight value, ranging from 0.2 to 1.0; the basic weight of deep semantic features in the no-fly zone is set to 1.0, and the basic weight of the original signal compression feature is set to 0.6; the basic weight of the original signal compression feature in the general surveillance zone is set to 0.8, and the basic weight of the deep semantic feature is set to 0.3.

[0030] When a flight is in the approach or takeoff phase, or when a suspected target crosses the runway boundary, the cloud sends a situation weight adjustment instruction to the corresponding detection node through the control instruction channel. This instruction increases the weight coefficients of deep semantic features and primary target candidate features, with the adjustment not exceeding 30% of the base weight. The final decision weight used for feature level selection is obtained by multiplying the base weight and the situation weight coefficient, and then clamping it within the range of 0.2 to 1.0. This weight serves as the core judgment parameter for feature level and coding rate selection.

[0031] Combining the current real-time network bandwidth conditions of the detection node, the control level of its grid-based control unit, and the real-time control situation weight, the decision weight is used as a correction coefficient for the bandwidth threshold to adaptively select the feature level and coding rate to be transmitted. In this embodiment, a three-level bandwidth benchmark judgment threshold is set. For every 0.2 increase in the decision weight, the bandwidth judgment threshold for the corresponding high-level feature is lowered by 1 Mbps: when the corrected available bandwidth is not less than 8 Mbps and the end-to-end latency is less than 20 milliseconds, the network conditions are considered excellent, and the transmission of deep semantic features is selected with 8 Mbps low compression ratio coding; when the corrected available bandwidth is between 2 Mbps and 8 Mbps, the network conditions are considered average, and the transmission of primary target candidate features is selected with 4 Mbps medium compression ratio coding; when the corrected available bandwidth is less than 2 Mbps and the packet loss rate is higher than 5%, the network conditions are considered limited, and the transmission of original signal compression features is selected with 1 Mbps high compression ratio coding.

[0032] In this embodiment, three coding rate levels are set under the same feature level. The code rate level is dynamically adjusted according to the packet loss rate and latency jitter parameters of the real-time network. When the packet loss rate increases by more than 2% or the latency jitter increases by more than 5 milliseconds in a single period, the code rate is switched down by one level. When the network parameters remain stable for three consecutive sampling periods, the code rate is switched up by one level.

[0033] When network link bandwidth is sufficient, latency and packet loss rate meet requirements, and the control level is high, deep semantic features are transmitted; when network link bandwidth is limited, latency jitter is large, and the control level is low, the original signal compression features with a high compression ratio are transmitted, and the deep feature extraction and target recognition calculation are completed by the cloud-based control processing platform.

[0034] The cloud-based control and processing platform receives feature data transmitted by detection nodes within the same gridded control unit. Based on the timestamp of the unified time reference of civil aviation air traffic control, it first performs joint calculation of spatiotemporal parameters such as arrival time difference, arrival angle difference, and Doppler frequency shift on the feature data of different detection nodes to complete the spatiotemporal alignment of the detection node features, and then carries out multimodal fusion processing of feature dimensions.

[0035] In this embodiment, multimodal fusion is achieved using a dual-branch cross-modal attention fusion network. The input consists of three types of feature vectors: radio frequency, acoustic, and optoelectronic, which are spatiotemporally aligned. By allocating channel attention and spatial attention weights between modes, the complementary feature contribution of weak signal modes is enhanced, and a fusion target feature with a unified 256-dimensional output is generated.

[0036] The target association matrix takes the fused target features and spatiotemporal solution parameters as input, and uses the feature cosine similarity, trajectory continuity and smoothness, and speed rationality of the same target among multiple nodes as association measurement dimensions. It combines the corresponding airspace altitude limit, low-altitude flight speed limit, and runway end no-fly boundary in the airport airspace control rules as hard constraints. The Hungarian algorithm is used to complete multi-node target matching and association degree scoring. In this embodiment, the association degree judgment threshold is set to 0.75. Candidate targets with an association degree lower than the threshold are judged as false targets and eliminated.

[0037] In this embodiment, each detection node maintains synchronization with the unified time reference of air traffic control through the BeiDou and GPS dual-mode timing module. The time synchronization accuracy is controlled at the level of 100 nanoseconds. Each segment of sensing data is marked with a unified UTC timestamp at the sampling start point, with a timestamp resolution of 100 nanoseconds.

[0038] During the calculation process, the sampling timestamps carried by the feature data are used as a reference. The peak time points of the same target feature at different nodes are determined by feature cross-correlation matching, and the arrival time difference is calculated. The incident angle of the target signal is corrected by combining the field orientation calibration value of the 8-element radio frequency antenna array of each node, and the arrival angle difference is obtained. The Doppler frequency shift parameter of the signal is extracted synchronously to assist in verifying the radial motion velocity of the target. Finally, based on the parameters of at least 3 effective detection nodes, the triangulation calculation of the three-dimensional position of the target is completed by multi-point intersection.

[0039] During the multimodal fusion processing, verification is carried out simultaneously. Airport airspace control flight rules are used as the constraint dimension to construct the compliance verification logic of target flight behavior. Combining whether the target trajectory breaks through the runway end no-fly boundary and whether the flight speed and altitude comply with the flight rules of the corresponding controlled airspace, a target correlation matrix is ​​constructed together through spatiotemporal consistency constraints and airspace control rule constraints.

[0040] In this embodiment, the target association matrix generates three scores for each candidate target: multi-node feature matching degree, trajectory smoothness, and flight rule compliance. If the multi-node feature matching degree is less than 0.6 and the signal arrival angle has a specular reflection mapping relationship with the terminal building glass curtain wall and the apron metal roof, it is determined to be a multipath reflection false target. If the signal spectrum features match the feature library of co-frequency equipment such as airport DME rangefinders and VHF voice communication devices, and the flight trajectory does not conform to the dynamic characteristics of small UAVs, it is determined to be a co-frequency interference false target.

[0041] False targets generated by multipath reflection and co-frequency interference are eliminated, and the trajectory and signal feature labels of suspected UAV targets that conform to the civil aviation low-altitude control data format are output.

[0042] When the cloud-based control processing platform determines, through feature fusion, that a silent suspected target without radio frequency signals appears within the no-fly zone warning area at the airport runway end and shows a tendency to break through the runway end control boundary, it generates an active detection control command.

[0043] In this embodiment, the determination of a silent suspected target is based on a combination of acoustic and photoelectric mode sensing characteristics and the absence of radio frequency mode signals: when both acoustic and photoelectric modes detect target signals that conform to the flight characteristics of a multi-rotor UAV, and no valid UAV flight control radio frequency signal and image transmission signal are captured within a 3-second sliding window in the corresponding airspace, it is determined to be a silent suspected target; the trend of breaking through the control boundary is determined by fitting the motion trend of the target's trajectory for 20 consecutive frames. When the shortest distance between the fitted predicted trajectory and the runway end no-fly boundary is less than 300 meters of warning distance and the target is flying towards the no-fly boundary, it is determined that there is a breaking trend.

[0044] By using the time slot scheduling mechanism of the 5G private network, the transmission time slot resources of the control sector to which the corresponding grid control unit belongs are reconstructed. Some idle time slots originally used for issuing control commands are reused as dedicated transmission time slots for active detection commands and echo data. Active detection control commands are preferentially issued to the detection nodes in the corresponding grid control unit.

[0045] In this embodiment, the time slot scheduling adopts 5GTDD time-division duplex mode, and the duration of a single time slot is set to 1 millisecond. During the reconstruction process, the non-urgent historical archive data transmission time slots in the corresponding sector are temporarily suspended, and no less than 4 consecutive time slots are transferred as dedicated transmission channels for active probe commands and echo data. The idle time slot reuse adopts a preemptive scheduling mechanism, and the scheduling priority of active probe commands is higher than that of regular monitoring data, ensuring that the end-to-end command delivery delay is controlled within 10 milliseconds.

[0046] The system controls the detection nodes to transmit directional radio frequency excitation signals and acoustic detection pulses that comply with the electromagnetic control requirements of civil aviation airspace. At the same time, it reserves dedicated transmission time slots for control of three adjacent detection nodes for real-time transmission of active detection echo data. After the echo data is transmitted back to the cloud-based control processing platform in real time, it is incorporated into the feature fusion and spatiotemporal parameter calculation process of the corresponding silent target to supplement the target's radio frequency and acoustic reflection characteristics and support subsequent identity matching and risk assessment.

[0047] In this embodiment, the echo features are generated by the local signal preprocessing of the detection node through matched filtering and pulse accumulation. The features include echo delay, echo intensity, and spectral distribution characteristics. After being transmitted back to the cloud, the echo features are input into the multimodal feature fusion network and spliced ​​with the original acoustic and photoelectric sensing features in terms of feature dimensions. The feature vector and confidence of the corresponding target in the target association matrix are updated synchronously, so as to realize the supplementary identification of silent targets from passive sensing to active reflection features.

[0048] In this embodiment, multiple detection nodes within the same control unit transmit excitation signals in a staggered manner according to a time slot scheduling allocation. The transmission timing interval is set to 2 milliseconds, which is greater than the sum of the maximum echo propagation time and pulse width of the excitation pulse, thus avoiding mutual interference and self-transmission / reception interference among multiple nodes. The radio frequency excitation signal uses a frequency point within the 5.8 GHz ISM open band, avoiding the dedicated frequency band for aviation communications, while the acoustic detection pulse uses the 40 kHz ultrasonic band. The transmission direction of the excitation signal is controlled by the cloud based on the estimated azimuth of the silent suspected target, and beam pointing is controlled by the 8-element phased array antenna of the node. The transmission power is dynamically set according to the airspace electromagnetic environment level, with the equivalent omnidirectional radiated power not exceeding 20 dBm, always meeting the electromagnetic radiation limit requirements for civil aviation airspace. During the transmission of the excitation signal, adjacent detection nodes synchronously switch to high-gain receiving mode to collect the echo signal reflected by the target.

[0049] The cloud-based control and processing platform will match and verify the calculated target trajectory and signal characteristics with the civil aviation drone control identity feature database. For targets with radio frequency signals, it will identify the device model, signal pattern, and control terminal feature identifier. For silent targets, it will identify the device model and flight behavior characteristics based on acoustic and photoelectric features, and match and confirm the target's compliance registration status and flight permissions.

[0050] In this embodiment, the matching verification adopts the cosine feature similarity comparison method, which compares the spectrum features and modulation mode features extracted from the target with the standard features of each model of UAV pre-stored in the identity database one by one. If the similarity is higher than 0.8, it is determined to be the corresponding model. The compliance registration status matching uses the unique equipment identification code of the UAV to perform dual association verification with the Civil Aviation Administration's UAV real-name registration database and the airport flight plan reporting system. Targets with matching registration information and within the airport reporting flight period are marked as compliant, and targets without matching registration information and without flight plan reporting are marked as unregistered.

[0051] Simultaneously, the same target captured by multiple detection nodes is subjected to dual verification of feature consistency and control behavior consistency to eliminate false targets with mismatched features or flight behavior that does not conform to control logic.

[0052] The feature consistency verification compares the deviations in the center frequency and modulation bandwidth parameters of the same target captured by different detection nodes. If the frequency deviation exceeds the allowable range of 50kHz, it is determined to be a false target. The control behavior consistency verification compares the target's flight trajectory and speed change trend with the flight rules of the corresponding airspace. If there are abrupt trajectories that do not conform to the flight dynamics characteristics of small UAVs, such as acceleration exceeding 15m / s² or climb rate exceeding 10m / s, it is determined to be a false target.

[0053] Output target detection results with control attribute tags, which are divided into four categories: airport inspection and bird control drones that have registered their flights at the airport are marked as compliant flights; drones that have been registered but have not reported their flight plans to the airport are marked as unregistered flights; drones without registration information that are moving toward the runway end no-fly zone are marked as suspected illegal flights; and drones with unknown signal characteristics and abnormal flight trajectories are marked as key targets of control.

[0054] The cloud-based control and processing platform generates parameters for an airport multipath interference adaptive detection model based on UAV target samples, multipath false detection samples, and airport electromagnetic interference samples collected in airport airspace control scenarios through targeted training.

[0055] In this embodiment, the model training adopts a two-step training process of basic model pre-training and airport scene sample fine-tuning: the first step is to complete the pre-training of the basic feature extraction network based on a general drone detection dataset containing 120,000 samples. The training batch is set to 32, the initial learning rate is set to 0.001, and the cross-entropy loss function and the mean squared error loss function are weighted in a 7:3 ratio as the total loss. After 100 rounds of iterative training, a general basic model is obtained. The second step involves fine-tuning the training using 20,000 UAV target samples, 15,000 terminal multipath reflection interference samples, and 10,000 negative samples of co-channel interference from airport communication equipment collected on-site at the airport. The learning rate during the fine-tuning phase is set to 0.0001, and the training is iterated for 30 rounds. During the training process, airport airspace boundary constraints and actual airport background electromagnetic noise fitting data are incorporated to finally obtain the airport multipath interference adaptation model parameters. The model parameters are then layered and packaged into convolutional layers, pooling layers, and fully connected layers, and then divided into transmission fragments according to a 1500-byte Ethernet MTU. After deducting the overhead of the IP protocol header and transport layer protocol header, the application layer data volume of a single transmission fragment does not exceed 1460 bytes, adapting to the 5G private network transmission capability.

[0056] Through the 5G private network's breakpoint resume and segmented transmission mechanism, the detection model parameters are segmented and distributed to the detection nodes within the airport's airspace coverage area in descending order of airspace control level.

[0057] In this embodiment, the model parameters are fragmented according to a 1500-byte Ethernet MTU. Each parameter fragment is accompanied by a unique consecutive sequence number and a 32-bit cyclic redundancy check code. After receiving the data, the probe node verifies each fragment. Fragments that fail verification are automatically retransmitted to the cloud, with the request carrying a list of the sequence numbers of the failed fragments. Resumption of transmission is based on the sequence numbers of the received fragments recorded locally by the probe node. After the transmission is interrupted and resumed, the probe node reports the largest sequence number that has been received to the cloud, and the cloud continues to send the data from the next fragment with the corresponding sequence number. During the transmission process, the transmission priority of control service data is always higher than that of model parameter fragment transmission to avoid model updates affecting normal probe services.

[0058] Without interrupting normal airport control detection operations, the detection nodes use a parameter hot-loading method to complete the online incremental update of the local detection model.

[0059] In this embodiment, the probe node locally sets up two independent model parameter running areas: a primary area and a backup area. The primary parameter area carries the model running for the current normal probe service, while the backup parameter area is used to cache model parameters to be updated. Newly received fragment parameters are first written to the backup parameter area. After all fragments are received and the overall verification is completed, the system first performs offline verification in the backup parameter area using 10 preset standard verification frames. After confirming that the model output confidence deviation is within 5%, the system selects a 100-millisecond interval between two adjacent 10Hz sensing data acquisition frames to perform parameter area switching. During the switching process, only the weight parameters of the convolutional layer, pooling layer, and fully connected layer are updated, without changing the model network structure. After the switching is completed, the new parameters are directly used to carry out subsequent probe services, and the currently acquired sampling frame data is not discarded throughout the switching process.

[0060] In this embodiment, the detection node locally stores the identification information of the current airport control scenario. When receiving model parameter fragments, it first checks whether the scenario identification matches. Only the parameter fragments that match are written into the spare parameter area. The node can store two sets of model parameters for different scenarios at the same time and call the corresponding parameter set according to the scenario switching instructions issued by the cloud.

[0061] The cloud-based control processing platform performs three levels of transmission scheduling for detection results based on the target's control attribute tags and corresponding control levels: control action level, control monitoring level, and control archiving level. Targets marked as suspected illegal flights (control action level) are directly transmitted to the airport tower control room and the airfield clearance command terminal via a dedicated low-latency control link. The corresponding target situation information is simultaneously pushed to the detection nodes in the three surrounding grids. The detection sampling frequency in the corresponding area is increased from 10Hz to 20Hz, the feature transmission level is raised by one level, and dedicated transmission bandwidth is allocated to enhance the control monitoring level. Targets marked as unregistered flights or those of key control concern (control monitoring level) are pushed to the airport operations command center control terminal via the control monitoring link. Targets marked as compliant flights (control archiving level) are transmitted back via a regular link and stored in the airport control database for archiving.

[0062] By receiving detection effect feedback data from detection nodes and handling result feedback data from tower control via 5G private network, the core optimization indicators are control handling success rate, number of control misoperations caused by false alarms, and control response latency for silent targets. The gridded control unit division, detection node topology, transmission resource configuration, and detection model weight of the corresponding control airspace are adjusted on a weekly basis.

[0063] In this embodiment, when the number of false alarms caused by a single grid in a single month exceeds 5, the side length of the grid unit in that area is reduced by 30%, and the number of detection nodes is added to the newly added grid unit according to the single grid node deployment standard. At the same time, the interference suppression weight of the detection model in the corresponding area is increased by 0.2 within the range of 0 to 1, and no further increase is made after reaching the upper limit. When the response delay from the detection of a silent target to the initiation of active detection exceeds 50 milliseconds, the dedicated transmission slice bandwidth of the corresponding sector is expanded to 20Mbps and the priority of active detection command scheduling is increased by one level. When the monthly control handling success rate is less than 90%, 500 new sample data of the corresponding scenario are added to iteratively fine-tune the detection model.

[0064] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for remote detection of a UAV based on network transmission, characterized in that, The specific steps include the following: According to the low-altitude air traffic control rules, the controlled airspace is divided into grid-based control units, and detection nodes are deployed in the grid-based control units. The network link quality parameters from each detection node to the cloud control processing platform are collected, and a transmission mapping topology is constructed by combining the control level of the control unit with the real-time control situation, and corresponding transmission links are allocated to each detection node. The system receives feature data uploaded by the detection node. The feature data is generated by the detection node through feature extraction of the collected sensing signals and bound to the control service weight. The detection node also adaptively selects the feature level and encoding rate for transmission based on the real-time network status and the control level it belongs to. Spatiotemporal parameters are jointly calculated and multimodal features are fused from the feature data of the detection nodes within the same gridded control unit. A target correlation matrix is ​​constructed in conjunction with airspace control flight rules. After removing false targets, the UAV target trajectory and feature labels are obtained. The cloud makes a judgment based on the target trajectory and feature label output by the multimodal fusion. When a silent suspected target without a corresponding radio frequency signal appears within the airspace warning range, an active detection and control command is generated. The transmission time slot resources of the corresponding control sector are reconstructed through the time slot scheduling mechanism, and the detection node is controlled to transmit the excitation signal. After receiving the echo data, the detection node transmits it back to the cloud and supplements the echo data into the feature fusion and spatiotemporal parameter joint solution process. The calculated target data is matched and verified against the UAV control identity feature database to complete the consistency verification of multi-node target features, and the UAV detection results with control attribute tags are output. 2.The network transmission based remote detection method of claim 1, wherein, The controlled airspace is divided into three levels: no-fly zone, restricted zone, and general surveillance zone; the grid-based control unit is a standardized unit, and each grid-based control unit is equipped with integrated radio frequency sensing, acoustic sensing, and photoelectric sensing detection nodes. 3.The network transmission based remote detection method of claim 2, wherein, The network link quality parameters include end-to-end latency, available bandwidth, packet loss rate, and latency jitter. Low-latency, high-reliability dedicated transmission slices are allocated to the high-level control units, while shared link scheduling is used for the control units corresponding to the general monitoring areas.

4. The network transmission based remote detection method of unmanned aerial vehicles according to claim 3, characterized in that, The feature data is divided into three levels: original signal compressed features, primary target candidate features, and deep semantic features; Each level of feature is pre-bound to the corresponding regulatory business weight, which corresponds to the regulatory-level signal tracing, early warning screening, and target identity verification requirements, respectively.

5. The network transmission based remote detection method of claim 4, wherein, The detection node selects the feature level and the encoding rate to be transmitted based on its own real-time network bandwidth conditions, the control level of the control unit to which it belongs, and the real-time control situation weight. When network conditions are favorable and the control level is high, deep semantic features are transmitted; when network conditions are limited and the control level is low, raw signal compression features with a high compression ratio are transmitted.

6. The method for remote detection of unmanned aerial vehicles based on network transmission according to claim 5, characterized in that, The spatiotemporal parameter joint solution is based on the timestamp of the unified time reference of air traffic control to complete the spatiotemporal alignment of the characteristics of different detection nodes. The solution parameters include time difference of arrival, angle difference of arrival, and Doppler frequency shift. The target correlation matrix is ​​constructed by combining spatiotemporal consistency constraints and airspace control flight rule constraints, and is used to eliminate false targets generated by multipath reflection and co-frequency interference.

7. The method for remote detection of unmanned aerial vehicles based on network transmission according to claim 6, characterized in that, The active detection control command is generated by the cloud when it determines that a silent suspected target has a tendency to break through the control boundary; the cloud uses a time slot scheduling mechanism to reuse idle time slots to issue detection commands and reserves dedicated return time slots; the excitation signal includes a directional radio frequency excitation signal and an acoustic detection pulse that meet the requirements of airspace electromagnetic control.

8. The method for remote detection of unmanned aerial vehicles based on network transmission according to claim 7, characterized in that, The matching verification is used to identify the target's equipment model, signal pattern, and control terminal characteristics, and to match and confirm the target's compliance registration status and flight permissions; the consistency verification includes dual verification of characteristic consistency and control behavior consistency; the output detection results are labeled with control attributes and are divided into four categories: compliant flight, unregistered flight, suspected illegal flight, and key control focus.

9. The method for remote detection of unmanned aerial vehicles based on network transmission according to claim 8, characterized in that, The detection model parameters, which are pre-trained based on scene samples, are distributed to the corresponding detection nodes through a segmented transmission mechanism, thereby controlling the detection nodes to complete the online update of their local detection models. The detection model parameters are obtained by targeted training of UAV target samples, false detection samples and interference samples in a dedicated control scenario. The detection model is classified according to the control scenario, including three categories: airport multipath interference adaptation model, urban dense target adaptation model and border silent target adaptation model. The model parameters are distributed from high to low according to the control level through the breakpoint resume and fragmented transmission mechanism; the detection node completes online incremental updates by using parameter hot loading.

10. The method for remote detection of unmanned aerial vehicles based on network transmission according to claim 9, characterized in that, Transmission scheduling is carried out according to the control level corresponding to the target, and feedback data on detection effects and control handling results are received to adjust the gridded unit division of the controlled airspace, the topology of detection nodes, the configuration of transmission resources and the weight of detection models. The transmission scheduling is divided into three levels: control and processing level, control and monitoring level, and control and archiving level. The control and disposal level targets are directly transmitted to the control and command terminal via a low-latency dedicated link; the control and monitoring level targets are pushed to the control terminal via a monitoring link; and the control and archiving level targets are transmitted back and archived via a regular link. The feedback data includes the detection effect feedback of the detection node and the handling result feedback of the control and command terminal. The cloud uses the control handling success rate, the number of control misoperations caused by false alarms, and the control response delay of the silent target as the core optimization indicators to perform corresponding adjustments to the gridded unit division of the control airspace, the topology of the detection node, the configuration of the transmission resources, and the weight of the detection model.