Joint optimization method for drone navigation error implantation and target capture

Abstract

This invention provides a joint optimization method for UAV navigation error implantation and target acquisition, comprising: collecting UAV data and unifying its time reference; obtaining the embedding vector and initial parameters of the navigation field through a spatial environment model; based on the initial parameters of the navigation field and the embedding vector, combined with the real-time state of the target UAV, synchronously transmitting illusion signals to implant navigation errors into the target UAV; updating the navigation field parameters using a joint optimization algorithm and continuously transmitting illusion signals to guide the target UAV to a predetermined acquisition position; determining whether to activate the acquisition device to perform deceleration acquisition based on the navigation field parameters and the real-time state of the target UAV; acquiring the net-landing attitude data, comparing it with the embedding vector to generate an update signal, and adjusting the embedding vector and the joint optimization algorithm. The beneficial effects of this invention are: it achieves simultaneous implantation of progressive navigation errors in three channels—airspeed measurement, vision, and magnetic compass—significantly shortening the guidance time and achieving low-energy flexible acquisition.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) management and capture, and in particular relates to a joint optimization method for UAV navigation error implantation and target capture. Background Technology

[0002] Drones are rapidly gaining popularity in logistics, high-altitude surveying, and public safety, but the risks of unauthorized flights and loss of control are also increasing. Drone capture solutions can forcibly take over the flight path without damaging the drone and guide it to a safe area, while also facilitating evidence collection and reuse. This has significant industry value for low-altitude safety management and emergency response.

[0003] Traditional capture methods rely on: ① radio frequency interference to forcibly break the link, followed by shooting down the drone with a mechanical net or interceptor missile; ② deceiving the drone with a single global satellite navigation system to change its course; ③ high-power directed energy weapons to burn out electronic components. The former method leaves the drone carrying residual kinetic energy after interference, making it vulnerable to injury upon crashing; the second method only alters the position estimate and becomes ineffective against multi-source fusion flight control systems; the third method causes significant electromagnetic pollution in urban areas and carries high legal risks, making it unsuitable for deployment in busy regions. Summary of the Invention

[0004] In view of this, the present invention aims to propose a joint optimization method for UAV navigation error implantation and target acquisition, in order to solve at least one of the above-mentioned technical problems.

[0005] To achieve the above objectives, the technical solution of the present invention is implemented as follows: The first aspect of this invention proposes a joint optimization method for UAV navigation error implantation and target acquisition, characterized by comprising the following steps: Collect UAV data and unify its time reference; obtain the embedding vector and initial parameters of the navigation field through the spatial environment model. Based on the initial parameters and embedding vector of the navigation field, combined with the real-time status of the target UAV, an illusion signal is synchronously emitted to implant navigation error into the target UAV; The navigation field parameters are updated using a joint optimization algorithm, and illusion signals are continuously emitted to guide the target UAV to the predetermined capture position. Based on the navigation field parameters and the real-time status of the target UAV, determine whether to activate the capture device to perform deceleration capture; Acquire netting attitude data, compare with the embedded vector to generate update signals, and adjust the embedded vector and joint optimization algorithm.

[0006] Furthermore, the process of constructing the space environment model includes: A three-dimensional coordinate system is generated by reconstructing three-dimensional point clouds, meshing, and registering optical and wireless signal data, and an embedding vector is established within the three-dimensional coordinate system using the navigation field node positions as indices.

[0007] Furthermore, the illusion signal includes: Acoustic illusion signal, wherein the acoustic illusion signal is emitted by a phased-array ultrasonic directional sound pressure field; Optical illusion signal, wherein the optical illusion signal is generated by projecting environmental model feature patterns from a digital micromirror array; A magnetic field illusion signal, wherein the magnetic field illusion signal is generated by an alternating magnetic field by a dual-coil device.

[0008] Furthermore, the timing of the transmission of the illusion signal is synchronized with a unified time reference to ensure that the navigation error remains consistent in time across the target UAV's airspeed channel, visual channel, and magnetic compass channel.

[0009] Furthermore, the joint optimization algorithm includes: Annealing search, which generates a set of candidate navigation field parameters; Reinforcement learning, wherein the reinforcement learning continuously adjusts the weights of the candidate navigation field parameter set based on the real-time residual between the predicted state information of the target UAV and the navigation error information; The annealing search and reinforcement learning are performed sequentially.

[0010] Furthermore, the activation of the capture device is based on the following determination: The spatial error of the navigation field parameter output is no greater than the preset spatial threshold. The speed error of the navigation field parameter output is not greater than the preset speed threshold; If both the space threshold and the speed threshold are met, a capture trigger signal is issued to start the capture device to perform deceleration capture.

[0011] Furthermore, the capturing device includes: An electrostatic adsorption layer, wherein the electrostatic adsorption layer is composed of an alternating polarity partitioned conductive fiber mesh. An electromagnetic eddy current braking layer is provided, wherein multiple turns of wire are embedded in a flexible substrate and eddy currents are generated by AC excitation to apply a reverse braking torque to the rotor of a target UAV.

[0012] Furthermore, the net-landing attitude data is encoded into an attitude vector by a neural network, and a cosine similarity calculation is performed between the vector and the embedded vector to generate an update signal.

[0013] Furthermore, the update signal is written into the distributed ledger after capture is completed, and is loaded as the embedding vector and initial weights of the joint optimization algorithm during the initialization of the next capture task.

[0014] The second aspect of this invention proposes a joint optimization system for UAV navigation error implantation and target acquisition, applied to the method described in the first aspect of this invention, comprising: The environmental acquisition module is used to collect optical data and wireless signal data and generate a unified time reference through a distributed consensus mechanism. The model generation module is used to construct a spatial environment model under a unified time reference to obtain the embedding vector and initial parameters of the navigation field; The illusion emission module is used to synchronously emit illusion signals based on the embedded vector, the real-time status of the target UAV, and the initial parameters of the navigation field, thereby implanting navigation errors in the airspeed channel, visual channel, and magnetic compass channel of the target UAV. The optimization and update module is used to use a joint optimization algorithm to couple the predicted state of the target UAV with the navigation error to update the navigation field parameters, and use the updated navigation field parameters to continuously drive the illusion emission module to emit illusion signals to guide the target UAV to the predetermined capture position. The capture execution module is used to determine, based on the navigation field parameters and the real-time status of the target UAV, that the target UAV is located in a predetermined space and its flight speed does not exceed a predetermined speed threshold. When this determination is made, the capture device, which includes an electrostatic adsorption layer and an electromagnetic eddy current braking layer, is activated to perform deceleration capture and outputs the net-landing attitude data. The feedback update module is used to compare the net-landing attitude data with the embedding vector to generate an update signal, and adjust the embedding vector and joint optimization algorithm parameters according to the update signal for subsequent capture.

[0015] Compared with existing technologies, the joint optimization method for UAV navigation error implantation and target acquisition described in this invention has the following advantages: By employing multimodal illusion signal (acoustic, optical, magnetic field) techniques, progressive navigation errors were simultaneously implanted into the three channels of airspeed measurement, vision, and magnetic compass, bypassing multi-source fusion verification.

[0016] By employing a joint optimization algorithm of annealing and reinforcement learning, millisecond-level adaptive adjustment of navigation field weights was achieved, significantly shortening the guidance time.

[0017] By employing a combined approach of electrostatic adsorption layer and electromagnetic eddy current braking layer, low-energy flexible capture is achieved, preventing objects from falling from the air.

[0018] By using posture feedback to write into the distributed ledger technology, a capture-learning closed loop is achieved, and the system evolves automatically with the task. Attached Figure Description

[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic flowchart of the method of the present invention; Figure 2This is a schematic diagram of the interaction of the system of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0021] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0022] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0023] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] like Figure 1 As shown, the joint optimization method for UAV navigation error implantation and target acquisition includes the following steps: S1. Collect UAV data and unify its time reference, and obtain the embedding vector and initial parameters of the navigation field through the spatial environment model.

[0025] In the scenario of deceiving and capturing drones, the implantation of any navigation error and path guidance depend on the spatiotemporal consistency of the external environment. Therefore, the airspace is first "multimodal integrated characterization" and a unique reliable time reference is formed with the support of a distributed consistency mechanism. Then, the embedding vector and the initial parameters of the navigation field are generated. The two together serve as the primary priors for subsequent illusion signal generation and integrated optimization.

[0026] The optical data section employs visible light and short-wave infrared fusion imaging. The visible light channel acquires texture and structure, while the short-wave infrared channel provides information on material and temperature differences. Both are resampled at the same scene and resolution to generate a 4D spectral cube. The wireless signal data section focuses on acquiring mid-frequency samples and broadband radio frequency background power spectra from global satellite navigation systems to determine satellite visibility, electromagnetic interference, and noise-terrain coupling characteristics. The mid-frequency samples directly correlate with the signal injection location for navigation decoys, while the radio frequency power spectrum reflects the on-site propagation characteristics. Both types of data are synchronized using a unified time reference to avoid coordinate drift caused by microsecond-level jitter.

[0027] The generation of the unified time base adopts a lightweight Byzantine fault-tolerant consensus algorithm. Each navigation field node and accompanying device broadcasts its local time scale in rounds of 250 milliseconds within the local high-speed link, and then confirms the maximum time scale through voting in the same round. The final result is written in the form of a block hash, carrying only an 8-byte timestamp and a 2-byte checksum, which ensures traceability and controls overhead. This timestamp is attached to the header of each subsequent frame of optical data and wireless signal data to achieve cross-modal synchronization.

[0028] The spatial environment model uses a 3D point cloud mesh hybrid to perform sparse stereo reconstruction on multi-band images to obtain an initial point cloud. The propagation loss is estimated using wireless signal data as the weight of each node in the point cloud. The weight and position form a 6-dimensional vector, which is then compressed to a fixed resolution mesh through voxelization. This grid is denoted as: ; in, Indicates the first 3D coordinates of the voxel center This represents the weight of the corresponding wireless signal propagation loss, and this expression serves as the main basis for the spatial distribution and power scheduling of subsequent illusion signals.

[0029] Using graph neural network mapping, Each node in the graph is treated as a vertex, and edges are formed between node pairs whose Euclidean distance is less than a set threshold. After three layers of graph convolution, a 156-dimensional real vector is output. ; in, For graph convolutional network parameters, Called the first Each embedding vector contains local terrain undulations, electromagnetic attenuation, and optical texture decoding results; The embedded vectors are used as indices for quickly retrieving the optimal phase and power configuration in subsequent acoustic illusion signal array shaping, optical projection texture selection, and magnetic field phase adjustment. The graph neural network training can be completed offline during system idle periods, and only weights are loaded during practical applications, ensuring that the inference latency is within 10 milliseconds.

[0030] The initial parameters of the navigation field are calculated jointly by the set of embedding vectors and the center position of the capture region. Assume the coordinates of the capture center are... The initial potential function is: ; in, For the first The weight coefficients of each navigation field node. in accordance with and The cosine correlation is evenly distributed. For any spatial point coordinates, this potential function serves as the initial state of the navigation gravity well in the subsequent joint optimization algorithm. By adjusting... Fine-grained shaping of the potential gradient enables incremental injection of navigation errors.

[0031] The environment model constructed using both optical and wireless signal channels preserves spatial geometry and reflects electromagnetic propagation characteristics, ensuring that the navigation field potential function remains consistent in both physical and electromagnetic spaces. A unified time reference ensures that multimodal data is aligned at the microsecond level, avoiding time-series drift in optical reconstruction and radio frequency analysis. Node feature embedding compresses complex environmental information into a fixed-length vector, significantly reducing the search dimension of the joint optimization algorithm. The initial parameters of the potential function ensure that the navigation error conforms to spatial constraints from the start, shortening the optimization convergence process.

[0032] Tests show that, compared with a static uniform weight scheme that does not use node embedding, the present invention reduces the initial yaw time of the target UAV by about 35% under the same interference power budget.

[0033] Based on the above scheme, the following example is given: In a field demonstration, three fixed navigation field nodes and one escort device were deployed. The optical channel resolution was 4000 x 2000 pixels, and five frames were acquired per second. The wireless channel used a general-purpose software radio peripheral with a sampling rate of 20 MHz. The synchronization delay of the distributed consensus network was less than 3 milliseconds. After generating a time reference, a point cloud of approximately 800,000 nodes was constructed. After voxelization, the number of nodes was reduced to 100,000. Feature embedding was then generated. After the potential function was initialized, the target UAV was guided into the capture area within 20 seconds. After the capture device was triggered, the capture was completed in 6 seconds. Compared with the scheme without wireless signal weighting, it would take 35 seconds to guide the target into the capture area in the same scenario. This shows that the present invention significantly improves the guidance efficiency and reduces energy consumption.

[0034] Preferably, the spatial environment model generates a three-dimensional coordinate system by reconstructing three-dimensional point clouds, meshing, and registering coordinates from optical data and wireless signal data, and establishes an embedding vector within the three-dimensional coordinate system using the navigation field node positions as indices.

[0035] In this invention, optical data and wireless signal data play the role of "dual-mode coparametric": the former provides geometry and texture, while the latter provides electromagnetic propagation and material reflection information. In order to generate predictable and controllable navigation errors in the navigation deception process, it is necessary to first map the two modal data to the same three-dimensional coordinate system, and then assign each navigation field node a vector description containing optical-electromagnetic dual attributes. The core implementation process is divided into three stages: point cloud reconstruction, coordinate registration, and feature embedding generation.

[0036] First, a 3D point cloud reconstruction is performed. Optical data is obtained using binocular or multi-view vision. With the baseline length and intrinsic and extrinsic parameters known, the depth values ​​of sparse feature points are obtained through disparity to obtain the initial optical point cloud. At the synchronization time, the wireless signal data outputs the ray direction through the angle of arrival estimation method (such as beamforming), and intersects with the known transmitter position to obtain the radio frequency pseudo point cloud. The two sets of point clouds share a unified timestamp within the same sampling period to avoid time drift. To reduce the uneven distribution of sparse feature points, this invention performs voxel downsampling on the optical point cloud, making the voxel side length consistent with the predetermined minimum path variable unit, thereby ensuring that each grid cell contains at least one valid point during subsequent grid division.

[0037] The coordinate registration process is based on the assumption of rigid transformation, let Indicating the first point in the initial optical point cloud The three-dimensional coordinates of the points Indicates the coordinates after registration. Represents the rotation matrix. Let represent the translation vector, then we have: ;in, For the number of points in the point cloud, the rotation matrix With translation vector The initial value is provided by the GPS positioning system itself and is estimated using an iterative nearest neighbor algorithm. The algorithm converges within 5 iterations. The registered optical point cloud and radio frequency pseudo-point cloud are fused according to the nearest neighbor criterion. If the spatial distance between the two is less than a threshold, the nodes are merged and attributes are added to the nodes. This property is calculated from the wireless signal propagation loss and is used to reflect the limitations of the electromagnetic environment on the navigation spoofing power budget.

[0038] After registration is completed, mesh generation is performed, dividing the 3D coordinate system into uniformly sized cubic voxels, with the center coordinates of each voxel marked. For all points falling into the same voxel, their geometric centroids are calculated as voxel coordinates, and the signal loss properties are weighted averaged. Voxelization can effectively reduce the computational complexity of subsequent graph convolutional networks and ensure uniform distribution of navigation field nodes. The voxel list is the candidate set of navigation field nodes. In the feature embedding generation stage, the voxel list is treated as vertices in a graph. Undirected edges are created for pairs of vertices whose Euclidean distance between them is less than twice the voxel's side length. Vertex attributes consist of three parts: 3D coordinates, optical texture gradient, and signal loss. The optical texture gradient is derived from the local gradient average of a 4D spectral cube within the voxel and can distinguish between building edges and homogeneous regions. The signal loss reflects the degree of occlusion of the voxel in the radio frequency domain. Vertex attributes are denoted as a six-dimensional vector. Mapping function through a three-layer graph convolutional network Generate embedding vectors: ; in, A 156-dimensional floating-point vector, representing the first... The comprehensive characteristics of each navigation field node in the geometric, electromagnetic, and textural domains serve as a unique index in subsequent acoustic array phase retrieval, optical projection texture selection, magnetic field phase adjustment, and weight update, thereby ensuring that the three types of illusion signals remain spatially coordinated.

[0039] In navigation deception applications, feature embedding serves the following purposes: First, it acts as a reference for the weights of the potential function. During system initialization, the cosine correlation is calculated based on the capture center coordinates and the node feature embeddings, and this correlation is mapped to weight coefficients. Secondly, during the joint optimization algorithm iteration, it is used to quickly locate the steepest gradient direction, avoiding the search falling into optical texture repetition areas or radio frequency blind spots. Thirdly, after capture, by comparing the netting pose vector and the node feature embedding vector through cosine similarity, it is possible to quickly determine which nodes contributed the most during the deception process, providing prior information for the next round of initialization.

[0040] Based on the above scheme, the following examples are given: This invention is compared in typical urban scenarios and open desert scenarios. Uniform weighting schemes without feature embedding require higher acoustic power in urban scenarios to overcome multipath effects. The present invention automatically reduces the weight of highly obstructed areas by leveraging signal loss properties, reducing the acoustic power budget by approximately 30% and shortening the initial decoy time by approximately 40%. In open desert scenarios, due to limited texture information, traditional methods are prone to positioning drift. This invention compensates for insufficient texture by using radio frequency loss properties, reducing the standard deviation of positioning drift by approximately 45% and significantly improving overall guidance accuracy. In a field test, three fixed navigation field nodes and one accompanying device were deployed in an airspace with a radius of 2 kilometers. The optical channel frame rate was 5 frames per second, with a spatial resolution of 0.3 meters; the radio frequency sampling rate was 20 MHz, and the bandwidth was 50 MHz. A point cloud of 800,000 nodes was obtained through registration, which was then reduced to 100,000 nodes through voxelization; graph convolution inference took 8 milliseconds. After constructing the potential function, five acquisition experiments were conducted on the same model of UAV, with an average guidance time of 19 seconds and a 100% acquisition success rate, showing a significant performance improvement compared to the control scheme.

[0041] S2. Based on the initial parameters of the navigation field and the embedded vector, combined with the real-time status of the target UAV, an illusion signal is synchronously emitted to implant navigation errors into the target UAV.

[0042] This invention employs a "multi-channel synchronous illusion injection" strategy in the navigation error implantation stage. The core idea is to map the spatial gradient contained in the embedded vector into illusion signals of three physical domains, and maintain the phase consistency of the three types of signals with a unified time reference. This enables the airspeed measurement channel, visual channel, and magnetic compass channel to generate error vectors with controllable amplitude and consistent direction, thereby allowing the target UAV's fusion filter to continuously output correction commands in the same direction as the potential gradient, creating conditions for subsequent trajectory convergence.

[0043] First, obtain the embedded vector set and the real-time status information of the target UAV, including the position vector. With attitude vector ,in Let the coordinates of the target UAV in the three-dimensional coordinate system be: Given the Euler angles of the aircraft in the airborne inertial navigation system, calculate the potential gradient from the target point to the acquisition center based on the initial parameters of the navigation field: ; in, Indicates the first The weight of each navigation field node, The node coordinates are represented by the potential gradient direction, which represents the direction of the heading correction required for the current deception. The gradient magnitude represents the correction magnitude. The acoustic illusion signal is generated by a directional sound pressure field from a ring ultrasonic array. Assume the array has a total of... The actuator, the first The emission phase of each actuator is The direction of the desired sound pressure is denoted as a unit vector. The commonly used phased array synthesis relation is adopted: ; in, For ultrasonic frequency, For the speed of sound, For the first The position vectors of the actuators in the array coordinate system, after being amplified by a hardware phase-shifting amplifier, form a steady-state node near the airspeed measurement probe, simulating minute changes in the incoming flow and generating airspeed drift. The amplitude of this drift is related to... Proportional, therefore no additional power iteration is needed, only weights Indirect adjustment.

[0044] The optical illusion signal is projected by a digital micromirror array onto a structured light pattern, projecting the potential gradient direction onto the camera's imaging plane. A set of sparse feature points located in this direction is selected, and the pixel displacement field is calculated based on the projection transformation. The pixel displacement size is also based on... Linear scaling is used to maintain the residual balance of the fusion filter, and the brightness of the structured light pattern is modulated by grayscale to achieve seamless integration with the ambient light and avoid triggering alarms by the visual anomaly detection module.

[0045] The magnetic field illusion signal generates an alternating magnetic field through a dual-coil circuit. Under a unified time reference calibration, the driving current employs phase modulation to ensure that the direction of the superimposed magnetic field is aligned with... Collinearity, peak coil current based on: ;in, To ensure that the magnetic compass deviation does not exceed the system monitoring threshold, high-frequency modulation is used to make the magnetic sensor a slow drift source without triggering the anti-magnetic interference processing procedure.

[0046] The three types of illusion signals need to be strictly synchronized. Within each frame control cycle, the system writes a unified time reference into the signal controller register and triggers hardware phase-locking. The acoustic array, digital micromirror and coil drive start outputting simultaneously, with jitter less than 20 microseconds. The fusion filter for the airspeed, visual, and magnetic compass channels receives error vectors with consistent direction and amplitude within the same time window. Since the internal consistency check cannot distinguish the signal source, the error is interpreted as inertial drift, resulting in the output of a heading adjustment control quantity. Based on the above scheme, the following example is given: In a semi-open airspace guidance test, the target UAV originally flew with a constant heading. After the system started, the potential gradient magnitude was calculated to be about 0.4 in the first second, generating a sound pressure field amplitude of 2 Pascal, a structured light displacement of 0.3 pixels, and a magnetic field superposition of 0.6 microtesla. The residual of the fusion filter remained within the random noise range and no anomaly was triggered. After 15 seconds, the target UAV deviated from the original route by 12 degrees and entered the search path of the capture area. The test group and the control group were compared and found that the scheme of the present invention can shorten the time of the first heading deviation by 38% while maintaining the same power limit. Another group of night tests showed that the optical illusion signal maintained the same error amplitude in the low light environment by increasing the pattern brightness, and the system has excellent robustness. The system deployed a 64-element ultrasonic array, a 1280x800 digital micromirror, and a dual-coil drive unit on a real site. The controller adopted a unified time reference for scheduling, and the hardware phase-locked jitter was less than 10 microseconds. The UAV entered the field from 3 kilometers away from the capture center and flew into a cylindrical area with a radius of 50 meters above the capture net after 20 seconds of continuous guidance. The total power budget was reduced by 25% compared with the traditional high-power single-frequency interference scheme, which verified the advantages of precise injection of synchronous illusion signal in terms of energy saving and concealment.

[0047] The synchronous illusion injection principle of this invention also supports expansion in multi-drone scenarios. It only requires maintaining an independent coefficient group for each target UAV in the potential gradient calculation, and multiple targets can be deceiving in parallel under the same hardware framework. This feature provides a feasible path for subsequent swarm UAV combat.

[0048] Preferably, the illusion signals include acoustic illusion signals, optical illusion signals, and magnetic field illusion signals. The acoustic illusion signals are emitted by a phased-array ultrasonic directional sound pressure field, the optical illusion signals are projected by a digital micromirror array to display characteristic patterns of an environmental model, and the magnetic field illusion signals are generated by a dual-coil device to produce an alternating magnetic field.

[0049] The navigation error implantation of this invention relies on three types of synchronous illusion signals: acoustic illusion signals, optical illusion signals, and magnetic field illusion signals. These three signals are triggered by a unified time reference within the same control cycle, ensuring that the airspeed measurement channel, visual channel, and magnetic compass channel receive error vectors with consistent amplitude and direction. This causes the multi-source fusion filter of the target UAV to misjudge, mistaking the potential gradient for inertial drift and automatically correcting its course.

[0050] Acoustic illusion signals utilize a ring-shaped phased-array ultrasonic array to generate a directional sound pressure field, in which the first... The position vector of each actuator is denoted as . The acoustic operating frequency is denoted as The speed of sound is denoted as The expected unit vector of the main lobe direction of the sound pressure is obtained by normalizing the potential gradient, denoted as . The actuator phase is passed through The calculation is performed, and the phase is written into the drive pulse via a delay line. ;in, For the first The emission phase of the actuator creates a quasi-static pressure difference near the pitot tube inlet of the target UAV, which is equivalent to introducing a small positive deviation into the airspeed probe. When the potential gradient magnitude increases, it is only necessary to increase the overall phase gradient slope without increasing the sound pressure amplitude, so the navigation error can be increased, and the power consumption is controllable.

[0051] Optical illusion signals are generated by projecting texture displacement onto the camera's imaging plane using a digital micromirror array. First, feature points generated from the spatial environment model are projected onto the image plane to obtain pixel coordinates. The potential gradient direction is projected as a vector onto the pixel plane. The system is based on the scaling factor. Calculate displacement The corresponding brightness texture is loaded onto the digital micromirror array, and the pixel displacement amplitude corresponds linearly to the potential gradient magnitude. The scaling factor is experimentally calibrated to ensure that the visual residual is within the random noise threshold. Since the digital micromirror array can be refreshed repeatedly at the kilohertz level, the optical illusion signal can rotate in real time following the target posture, maintaining the error direction consistent with the potential gradient.

[0052] The magnetic field illusion signal generates an alternating magnetic field through a spatially orthogonal double coil. Let the unit vector of the target magnetic field superposition direction be... The two coils are arranged along the orthogonal axis and synthesize the magnetic flux density in the target direction by bipolar drive. The magnetic field amplitude is controlled by the magnetic flux density. ,in, , where is the magnetic flux density This is the coil sensitivity calibration coefficient. The potential gradient magnitude is used because the magnetic compass response to high-frequency interference decreases sharply with frequency. This invention selects a carrier wave of tens of kilohertz and uses pulse width modulation to achieve amplitude control, so that the magnetic field is regarded as a low-frequency drift rather than an instantaneous pulse by the inertial navigation system, thus bypassing the inertial navigation reset logic.

[0053] The synchronization mechanism of the three types of illusion signals is based on a unified time base and uses the same frame number. The central controller broadcasts a timestamp to the acoustic, optical and magnetic field drivers at the beginning of each control cycle. The hardware phase-locked loop compresses the local clock deviation to less than 10 microseconds. Experiments show that if the time drift exceeds 50 microseconds, the three-channel error vector will exhibit an angled rebound, making it easy for the target UAV's fusion filter to detect. Therefore, strict synchronization is the key to achieving covert navigation deception.

[0054] Based on the above solution, the following example is given: Comparative tests were conducted in a typical urban canyon scenario. Traditional single-domain interference (sound only or magnetic field only) requires high power to deflect the drone's heading and is prone to triggering anomalies. This invention uses a three-channel synchronized illusion signal, reducing the drone's heading deflection time from 30 seconds to 18 seconds with a total power reduction of approximately 35%, without triggering navigation alarms in the flight control system. Under low-light conditions at night, by increasing the brightness of the optical texture, the optical illusion signal still maintains a 0.3 pixel-level displacement, demonstrating that the synchronized adjustment scheme has adaptive capabilities to ambient light. The array is configured with 64 ultrasonic transducers operating at 40 kHz; the digital micromirror imager has a resolution of 1280×800 pixels and a refresh rate of 360 frames per second; the dual coils have an inner diameter of 0.8 meters, 240 turns, and a drive frequency of 14 kHz. The target UAV weighs 2.5 kg and starts 3 km from the acquisition center. After system startup, the system guides its course to the acquisition area within 15 seconds and completes deceleration and acquisition within 20 seconds. Example 2: In an open desert environment targeting a fixed-wing target, the system automatically increases the acoustic power by 15% and decreases the magnetic field amplitude by 10% due to gusts of wind attenuation, yet still completes acquisition within 25 seconds, demonstrating that the amplitude adjustment of the illusion signal can adapt to environmental changes.

[0055] Preferably, the transmission timing of acoustic illusion signals, optical illusion signals, and magnetic field illusion signals is synchronized with a unified time reference to ensure that navigation errors remain consistent in time across the target UAV's airspeed channel, visual channel, and magnetic compass channel.

[0056] In navigation deception scenarios, the illusion signals must arrive simultaneously at the airspeed measurement channel, visual channel, and magnetic compass channel of the target UAV. Otherwise, the multi-source fusion filter will detect the anomaly through timestamp inconsistencies and trigger a backup navigation mode. This invention employs a unified time reference and phase synchronization mechanism to ensure that the acoustic, optical, and magnetic illusion signals are aligned within sub-millisecond errors, thereby allowing the three-channel error vectors to exhibit homogeneous asymptotic drift within the fusion filter. The working principle and application are explained below from four dimensions: time reference generation, hardware synchronization implementation, software scheduling logic, and effect evaluation. The unified time base originates from a distributed consensus mechanism. Each navigation field node and accompanying device has a built-in rubidium clock. The initial deviation between local time bases is typically on the order of microseconds. The system uses 250 milliseconds as a consensus round, and Byzantine fault-tolerant voting is jointly completed by time broadcast messages and threshold signatures to select the largest time stamp in that round, which is recorded as the global timestamp. The timestamp is written in the header of all communication data packets in the form of an 8-byte unsigned integer and distributed to the acoustic array, digital micromirror array and magnetic field driver via hardware pulses. To avoid clock drift, the node uses a phase-locked loop to fine-tune the local clock cycle by cycle in subsequent cycles to keep the cumulative error within 10 microseconds.

[0057] Hardware synchronization adopts a "common trigger pulse" method. The central controller outputs a narrow pulse every 30 milliseconds. The pulse is sent to the timing control modules of the three drivers simultaneously along the differential pair line. After receiving the trigger pulse, the acoustic array starts to refresh the phase register with the array phase table. The digital micromirror array starts a pattern flip after the trigger pulse, and the refresh period is fixed to an integer multiple of the trigger period. The magnetic field driver resets the carrier phase accumulator according to the trigger pulse to keep the phase shift of the alternating magnetic field period consistent with the transmission period of the acoustic array. The actual timing deviation of the three signals measured by the experiment does not exceed 18 microseconds, which meets the detection threshold requirements of the fusion filter.

[0058] The software scheduling logic ensures the synchronization of the amplitude and direction of the three signals. In each control cycle, the system first calculates the potential gradient vector based on the embedded vector and the real-time state of the target UAV. Normalize the vector to obtain the direction vector. This vector serves as both the direction of the acoustic array's target beam and the direction of optical pattern center offset and magnetic field superposition, along with the potential gradient magnitude. This is used to map linear coefficients to sound pressure level amplitude, pixel displacement amplitude, and magnetic field amplitude, respectively. For synchronization purposes, all amplitude calculations are completed 5 milliseconds before the trigger pulse arrives, and the calculation results are entered into each driver queue to ensure that the hardware outputs immediately upon receiving the pulse. To quantify the impact of synchronization error on the effectiveness of navigation deception, this invention introduces an upper bound formula for synchronization phase error: ;in, This indicates the maximum permissible phase difference among the three channels. Indicates the acoustic carrier frequency. This indicates the upper limit of timing deviation. Taking a 40 kHz acoustic carrier as an example, if... microseconds, upper limit of phase difference Approximately 4.5 degrees, less than the half-power angle of the sound beam, ensuring that the sound pressure peak and the optical pixel shift peak remain in the same direction, as shown in the formula. Consistent with the actual hardware; Obtained through experimental calibration, it represents the maximum time error of the signal arriving at the target sensor. This invention uses hardware phase-locked loop and periodic refresh to... By controlling the signals within a specified threshold, the three signals are made to appear as coherent signal sources at the target sensor level.

[0059] Based on the above scheme, the following example is given: The effect of the synchronization mechanism is verified by comparative experiments. The control scheme only uses an independent clock to trigger, and the actual timing deviation is about 120 microseconds. This causes the angle between the acoustic error vector and the optical error vector to remain above 15 degrees for a long time. The target UAV's fusion filter detects the inconsistency within 10 seconds and switches the inertial navigation mode, thus the deception fails. After adopting the synchronization scheme of this invention, the angle between the three signal errors is less than 4 degrees for a long time. The target UAV's flight trajectory deviates smoothly from the original path without triggering any alarms. The average first yaw time of the 5 groups of tests is shortened by about 35% compared with the control scheme.

[0060] The fixed-wing target flies at a speed of 15 meters per second, with 3 navigation field nodes. The system has a trigger cycle of 30 milliseconds, a 64-channel ultrasonic array, a digital micromirror array refresh rate of 330 frames per second, a magnetic field superposition frequency of 14 kHz, and a synchronization phase difference within an 8-microsecond fluctuation range. After the decoy begins, the target yaws to 12 degrees and enters the acquisition zone 14 seconds later. The effective acoustic intensity throughout the process is at the 90-decibel level, with minor adjustments to the optical projection brightness. The peak magnetic field strength is 0.7 microtesla, and power consumption is reduced by 28% compared to a asynchronous high-power single-channel solution. In the scenario of swarm drones, the synchronous trigger pulse is extended to a multicast mode. Each accompanying device independently calculates the potential gradient under a unified time reference and synchronously outputs three types of illusion signals while maintaining the consistency of the error direction of the three channels. Actual demonstration shows that when three drones enter the capture area at the same time, the system synchronization scheme can avoid signal crosstalk between different targets, verifying the scalability of multi-target systems.

[0061] S3. Update navigation field parameters using a joint optimization algorithm and continuously emit illusion signals to guide the target UAV to the predetermined capture position.

[0062] The core of the navigation deception phase lies in continuously adjusting the navigation field parameters to ensure that the potential gradient direction remains consistent with the correction command of the target UAV. This invention employs a joint optimization algorithm that combines annealing search and reinforcement learning in parallel to achieve this goal. Annealing search is used to quickly locate a set of candidate navigation field weights in a high-dimensional discrete space, while reinforcement learning is responsible for performing continuous fine-tuning near the candidate set to reduce error residuals and shorten convergence time. The algorithm is executed cyclically, and the output navigation field parameters are written into the acoustic array, digital micromirror array, and magnetic field driver in real time to achieve real-time closed-loop signal processing.

[0063] The extended Kalman filter provides the predicted state vector of the target UAV. ,in The system represents the target's position in a three-dimensional coordinate system, and simultaneously extracts a fused residual vector from airspeed, visual, and magnetic compass channels. The annealing search phase is performed on the navigation field node set. Define a primary energy function: ;in, For navigation field weights, To capture the center coordinates, As the residual tradeoff coefficient, annealing search obtains a set of candidate weights with the lowest local energy by accepting alternative weight vectors with decreasing probability as the temperature gradually decreases. Reinforcement learning then takes over the set of candidate weights, and the agent treats each set of weights as an action, using a negative energy function. As an immediate reward, the state is defined as follows: The action is The strategy employs proximal policy optimization. The policy network generates fine-grained increments centered on candidate weights, adjusting the weight distribution in a continuous space to maximize the reward at the expected discounted return. The learning rate and gradient clipping threshold are relatively large in the early stages of training to accelerate exploration; they are automatically reduced when the energy function's rate of decline falls below a set threshold to prevent excessive oscillation. The final weights are output from the two-stage results. And calculate the potential gradient vector: ,in, For the first The coordinates of each navigation field node are such that the potential gradient direction is the unified direction of the three types of illusion signals. The potential gradient magnitude is mapped to the sound pressure amplitude, pixel displacement amplitude, and magnetic field amplitude respectively through the scaling factor, thereby realizing the amplitude closed loop of navigation error.

[0064] Based on the above scheme, the following example is given: Multiple rounds of experiments show that the joint optimization algorithm can reduce the energy function by about 25% compared with the single annealing or single reinforcement learning method. When simulating a complex urban canyon environment, the traditional static weight scheme has an average guidance time of 32 seconds, while the scheme of this invention reduces it to 19 seconds, and the root mean square deviation of the track is reduced by 40%. This result shows that the algorithm can quickly find the optimal gradient direction under the conditions of multipath occlusion and signal fading, thus improving the deception efficiency. The experiment deployed three navigation field nodes, whose coordinates were calibrated using the Differential Global Navigation Satellite System. The initial temperature of the algorithm was set to 1, the temperature decay rate was 0.9, the weight vector was sampled 150 times per round, the hidden layer width of the reinforcement learning policy network was 128, and the discount factor was 0.95. After the UAV entered the decoy airspace, the system output the first weight update at the 3rd second, the energy function dropped to 50% of the initial value at the 10th second, and the angle between the potential gradient direction and the capture center was less than 5 degrees at the 18th second. The illusion signal guided the UAV to quickly turn towards the capture net. Throughout the process, the peak sound pressure, projection brightness, and magnetic field amplitude remained below the pre-set power threshold, and no self-test alarm was triggered. By dynamically coupling the predicted state and navigation error through a joint optimization algorithm, this invention achieves efficient, stable and covert navigation field parameter updates, enabling the deception process to converge quickly while keeping power consumption below regulatory thresholds, providing a replicable and scalable solution for UAV capture systems.

[0065] Preferably, the joint optimization algorithm sequentially executes an annealing search step and a reinforcement learning step, wherein the annealing search step generates a candidate navigation field parameter set, and the reinforcement learning step continuously adjusts the weights of the candidate navigation field parameter set based on the real-time residual between the target UAV's predicted state information and navigation error information. The joint optimization algorithm of this invention consists of an annealing search step and a reinforcement learning step in sequence. The two share a unified energy function and work together with a unified time reference to complete efficient weight adjustment. The annealing search explores globally in a high-dimensional discrete space, while the reinforcement learning performs continuous fine-tuning in the neighborhood of candidate solutions. The combination of the two enables the navigation field parameters to converge quickly and stably to the optimal region in a dynamic environment.

[0066] Extended Kalman filter generates the target UAV predicted position vector With multimodal navigation error vector The energy function is set to ;in, For navigation field node weights, To capture the center vector, The residual trade-off coefficient is... Represents the Euclidean norm, and the annealing search is based on the initial temperature. Perform discrete perturbation, randomly replacing several elements in the weight vector in each round. Accept if necessary, otherwise proceed by probability. Accept, new temperature as per After approximately 10 rounds of searching, the temperature drops to a threshold, and the 10 weight vectors with the lowest energy are output as the candidate set.

[0067] Reinforcement learning uses candidate weights as action prototypes, and the state vector is derived from... and The splicing action involves incremental weighting. The policy network adopts a three-layer fully connected structure, with hidden layers containing 128, 64, and 32 nodes respectively. The activation function is Swish, and the reward is defined as... The algorithm employs a near-end policy optimization mechanism with a pruning threshold of 0.2 and a discount factor of 0.95. Inference time is approximately 2 milliseconds, and training time does not exceed 6 milliseconds, meeting the 30-millisecond control cycle requirement. The final weights are obtained by adding the increment of the policy network output to the candidate weights. ; Calculate the potential gradient vector based on the gradient vector at the current position, where the gradient vector at the current position is... Used to indicate the direction of navigation error injection ,in, For the first The coordinates of each navigation field node To correspond to the weights, the potential gradient direction is used to unify the direction of the three types of illusion signals. The potential gradient magnitude is mapped to the sound pressure amplitude, pixel displacement amplitude and magnetic field amplitude respectively through the scaling factor to realize the error amplitude closed loop. Annealing search and reinforcement learning share the energy calculation and residual evaluation modules to avoid redundant calculation burden. The state reading is completed 5 milliseconds after the start of each control cycle, annealing is completed 10 milliseconds, reinforcement learning inference and policy update are completed 18 milliseconds, and the weights are written to the acoustic array, digital micromirror array and magnetic field driver 20 milliseconds. The weight vector is normalized before being written to ensure that the total power is less than the system budget limit.

[0068] Based on the above scheme, the following example is given: The joint optimization algorithm improves the energy convergence speed by about 2.8 times compared with the pure annealing scheme, and the first potential gradient correct alignment time is reduced from 300 milliseconds to 110 milliseconds, which fully demonstrates the advantages of candidate set plus continuous fine-tuning. In the actual test in the urban canyon environment, the system outputs a weight update every 30 milliseconds, and the energy can be reduced by 10% on average in 3 frames. The target UAV is guided to the center of the capture area within a radius of 50 meters within 19 seconds. Compared with the static weight scheme, it takes 32 seconds. The total power consumption is reduced by about 25%, and no flight control alarm is triggered. The system has 3 navigation field nodes, with a node power amplifier limit of 8 watts, an initial annealing temperature of 1, and 150 weight vectors sampled per round. The reinforcement learning batch size is 64. When the target UAV enters the field from 3 kilometers away, the system reduces its energy to 60% of the initial level in the 3rd second, to 30% in the 10th second, and completes the acquisition zone positioning in the 18th second. The peak sound pressure level is 85 dB, the peak magnetic field level is 0.7 μT, and the optical brightness change is less than 10 lux. Compared with the fixed weight scheme, the target could not be introduced into the acquisition zone within 30 seconds, thus verifying the performance advantage of the algorithm of this invention.

[0069] S4. Based on the navigation field parameters and the real-time status of the target UAV, determine whether to activate the capture device to perform deceleration capture.

[0070] After the target UAV enters the decoy field, the navigation field parameters reflect the spatial potential gradient in real time, while the state vector output by the fusion filter provides the current position, velocity and attitude estimation. This invention combines the two to construct a capture trigger arbiter, which considers the two conditions of "position satisfied" and "velocity satisfied" in turn. As long as either of the indicators has not been met, the capture device remains on standby to avoid premature deployment that would lead to energy waste or accidental damage to friendly aircraft.

[0071] The position determination uses the centroids of the three navigation field nodes with the highest weights in the navigation field parameters as the dynamic acquisition centers, and denotes the real-time position vector of the target UAV as... The dynamic capture center vector is denoted as ; Calculate spatial error ;in, For Euclidean distance, a predetermined spatial threshold is set. If and only if The position is considered to satisfy the condition, and the velocity is considered to satisfy the condition from the velocity vector output by the extended Kalman filter. Mold taking ; With the predetermined speed threshold In comparison, when The system assumes that the speed and position are both satisfied, and the result of the logical AND of the two Boolean values ​​is used as the capture trigger flag signal. Once the flag signal is true, the controller immediately sends a start command to the capture device. The capture device consists of an electrostatic adsorption layer and an electromagnetic eddy current braking layer stacked on top of each other. The surface of the electrostatic adsorption layer is covered with a conductive fiber mesh, with positive and negative electrodes alternately connected in a checkerboard pattern. A stable potential is obtained from a high-voltage power supply. The soft material substrate can be stretched into a tension surface by a servo mechanism upon receiving a start command. The adsorption principle is based on Coulomb force. ;in, To absorb traction force, Where is the dielectric constant. This represents the equivalent adsorption area. To apply voltage, As an equivalent gap between the plates, the mesh surface provides longitudinal traction before overcoming the residual kinetic energy, pressing the UAV body toward the center of the capture port.

[0072] The electromagnetic eddy current braking layer, located beneath the electrostatic layer, is a planar spiral coil composed of a flexible polyimide film embedded with multiple turns of copper wire. A high-frequency alternating current generates a changing magnetic field beneath the film. When the conductive part of the target UAV's rotor enters this magnetic field region, eddy currents are induced within the metal conductor according to Faraday's law of electromagnetic induction. The eddy current magnetic field is opposite in direction to the original magnetic field, forming a reverse braking torque, which is controlled by angular velocity. The instantaneous angular velocity of the UAV rotor, and the braking torque generated by the eddy current, can be approximated as: ,in, As a constant related to the number of coil turns, conductor resistance and magnetic field amplitude, the braking torque can be adjusted by adjusting the driving current amplitude without changing the structure, thus achieving adaptive braking of target UAVs with different masses and rotation speeds; The electrostatic adsorption layer and the electromagnetic eddy current braking layer work together: the adsorption layer provides axial traction, and the eddy current layer is responsible for reducing rotor momentum. The system first rapidly increases the output of the high-voltage power supply, so that the mesh surface reaches the target potential within tens of milliseconds. At the same time, the driver increases the AC current to the rated value. The traction force and braking torque work together to reduce the remaining kinetic energy of the UAV. The formula for calculating the remaining kinetic energy is as follows: ,in, For the quality of drones, As the equivalent rotational inertia of the rotor, experiments show that the braking layer can [reduce / deflect] within 2 seconds. Once the drone drops below the safety threshold, the traction force of the electrostatic adsorption layer ensures that the drone will not detach from the capture port again during this period. The attitude acquisition process is achieved by arranging a contact matrix around the capture port. The contact array has a resolution of 64×64, and each contact contains a piezoresistive element. The acquisition frequency is 1 kHz. The resistance change caused by the pressure on the contact is decoded into a contact pressure matrix. Then, the body attitude vector is obtained through centroid calculation. The attitude vector is compared with the embedded vector using cosine similarity. Nodes with large differences indicate that they have made a high contribution to the deception process. The system feeds this information back to the joint optimization algorithm to adjust the learning rate, so that nodes with better performance are given priority in the next round of tasks.

[0073] Based on the above scheme, the following example is given: This capture strategy has three significant effects. First, the position determination depends on the real-time navigation field parameters, and the dynamic capture center can adaptively shift with changes in the environment, ensuring the overall convergence of the guidance path. Second, the double-layer braking structure separates adsorption and deceleration, realizing universal capture of different types of UAVs, and avoiding hard structure collisions and debris damage. Third, the attitude feedback closed loop allows the algorithm weights to self-evolve, and the potential function converges faster after multiple tasks. The capture device was deployed in an open area in the suburbs. An alternating potential of 45 kV was applied to the electrostatic adsorption layer, and the electromagnetic braking layer was driven at a frequency of 14 kHz with 240 coil turns. The test target was a multi-rotor UAV with a rotor diameter of 0.4 m. At the time of triggering, the UAV's speed was 3.2 m / s and the rotor angular velocity was 230 rpm. Within 246 milliseconds, the traction force locked the UAV's center within 150 mm of the center of the net. At 420 milliseconds, the eddy current braking reduced the rotor angular velocity to 90 rpm. The overall kinetic energy dissipation rate exceeded 92%. The capture success rate was 100% after 5 repeated tests, which verified the effectiveness of the judgment mechanism, the dual-layer braking hardware, and the feedback optimization logic.

[0074] Preferably, the activation of the capture device is based on the following joint determination: The spatial error of the navigation field parameter output is no greater than the preset spatial threshold. The speed error of the navigation field parameter output is not greater than the preset speed threshold; When the joint determination is established, a capture trigger signal is issued to activate the electrostatic adsorption layer and the electromagnetic eddy current braking layer.

[0075] The guiding objective of the navigation deception phase is to gradually pull the target UAV into the effective working area of ​​the capture device while ensuring the concealed injection of navigation errors. In order to avoid energy waste due to premature capture or escape due to late capture, this invention proposes a dual threshold joint judgment mechanism of "spatial error - velocity error". All the quantitative indicators required for judgment are derived from the navigation field parameters and the real-time status of the target UAV. It does not rely on external ranging equipment and does not increase the burden on the communication link. Spatial error is measured by the Euclidean distance between the dynamic acquisition center in the navigation field parameters and the real-time position of the target UAV. The dynamic acquisition center is updated in real time according to the current navigation field node weights: the coordinate vectors of the three navigation field nodes with the largest weights are taken. , , Calculate the centroid ; Real-time position vector of target UAV The spatial error, defined by the output of the extended Kalman filter, is... ,in, This represents the straight-line distance from the target drone to the dynamic capture center, with a preset spatial threshold. The value is determined by both the aperture of the capture device and the electrostatic adsorption coverage area. If the position meets the capture requirements, then the position is considered to have met the capture requirements. and All are three-dimensional coordinate vectors; Represents the Euclidean norm; The velocity error uses the target UAV's current velocity vector. modulus As an evaluation indicator, a preset speed threshold is used. The combined constraint from the maximum traction force of the electrostatic adsorption layer and the deceleration capability of the electromagnetic eddy current braking layer, when This indicates that the target drone's kinetic energy is below the limit of the capture net's support, allowing for safe adsorption. Considering different drone models and environmental wind speeds, this invention will... Set the relative speed to 80% of the maximum that the capture device can withstand, to leave a safety margin.

[0076] Spatial error and velocity error constitute a Boolean criterion: ; When the trigger Boolean value is set to true, a capture trigger signal is generated. The central controller simultaneously sends a synchronization start command to both the electrostatic adsorption layer and the electromagnetic eddy current braking layer. The voltage generator begins to boost the voltage, and the coil driver starts the carrier generator. The two actuators maintain a timing deviation of less than 2 milliseconds to ensure that traction and deceleration work in tandem. The trigger signal is sent in the form of a one-bit pulse through a differential pair line at the hardware level. The pulse width is 1 microsecond and the actual line delay does not exceed 300 nanoseconds. The electrostatic adsorption layer uses a cascaded Martens generator module. The boost curve is controlled by a digital signal processor in a closed loop to make the boost slope synchronously match the predicted value of the remaining kinetic energy, which avoids instantaneous current peaks and provides sufficient traction. The electromagnetic eddy current braking layer adopts a full-bridge inverter topology. The drive phase is reset when the capture command arrives, so that the induced magnetic field is opposite to the rotor cutting direction, achieving rapid deceleration. The advantage of the dual-threshold determination principle lies in simultaneously constraining two key variables: position and velocity. If only the position threshold is set, the target UAV will pass by the capture device at high speed, resulting in insufficient traction. If only the velocity threshold is set, the target will decelerate outside the capture area but still cannot be attracted. The joint determination ensures that the target is both within spatial reach and in a low kinetic energy state at the moment of capture, thus improving the capture success rate.

[0077] Based on the above solution, the following example is given: The effectiveness was evaluated by comparing two environments: urban high-rise canyons and open deserts. When using single location determination, the capture success rate in urban scenes dropped to 78%, mainly due to strong gusts causing the drone to quickly cross the capture area. In desert scenes, single speed determination achieved an 83% success rate, with common failures occurring when the target decelerated but its position deviated from the capture net. After adopting joint determination, the capture success rate in both scenarios exceeded 98%, and the average capture time was reduced by approximately 22%. A capture device was set up at the test site, with an electrostatic adsorption layer potential of ±45 kV, an electromagnetic eddy current braking carrier frequency of 14 kHz, and a preset spatial threshold. The speed threshold is 0.5 meters. The target UAV has a speed of 0.3 meters per second, a mass of 2.3 kilograms, an initial velocity of 4 meters per second, and navigation field parameters output after 16 seconds of guidance sequence execution. rice, With a modulus of 0.28 m / s, and simultaneously meeting the dual thresholds, the electrostatic layer rises to the target potential within 170 milliseconds after the controller issues the trigger pulse, the electromagnetic braking coil reaches the rated current, and the rotor angular velocity drops from 220 radians per second to 75 radians per second within 2 seconds, reducing the residual kinetic energy by 92%, thus completing the capture. Throughout the entire process, the peak power is 27% lower than the test limit. If multiple targets enter the capture zone simultaneously, this invention maintains an independent weight vector for each UAV in the navigation field parameters, enabling each target to have an independent weight vector. and The judgment process is such that the capture device can be configured with multi-segment electrodes and segmented coils, activating only the segment aligned with the target to avoid mutual interference. In the test with three targets in the same field, the system successfully distinguished and captured all targets in sequence, verifying the multi-target expansion capability.

[0078] Preferably, the electrostatic adsorption layer of the capture device adopts an interactive polarity partitioned conductive fiber mesh, and the electromagnetic eddy current braking layer has multiple turns of wire embedded in a flexible substrate and generates eddy currents through alternating current excitation to apply a reverse braking torque to the rotor of the target UAV.

[0079] The capture device is deployed on the top or side wall of the capture area. Its main purpose is to achieve flexible adsorption and rapid deceleration without hard impact when the kinetic energy of the target UAV has dropped to a low value, thereby ensuring the integrity of the aircraft structure and the safety of personnel. The present invention adopts a double-layer stacked structure: the upper layer is an electrostatic adsorption layer and the lower layer is an electromagnetic eddy current braking layer. The two layers are integrated into a sheet through the same flexible substrate. The start-up sequence is triggered simultaneously by the capture trigger signal, and the timing deviation is less than 2 milliseconds. The substrate of the electrostatic adsorption layer is a polyimide fiber mesh with an overall thickness of approximately 0.3 mm. After a conductive graphene film is deposited onto the fiber surface using a sputter coating, the mesh is divided into non-short-circuit positive and negative electrode zones. The width of each zone is selected based on the typical width of the target UAV, generally ranging from 40 mm to 60 mm. Adjacent zones have opposite polarities, forming an alternating polarity pattern. When the high-voltage power supply is boosted to the operating potential, a strong electric field is established between adjacent zones. If two adjacent zones are considered as parallel electrodes, the traction force can be approximated by Coulomb force. ,in, The vacuum permittivity, For the equivalent plate area, To apply voltage, As an equivalent gap, since the shell of drones is mostly made of carbon fiber or metal materials, the conductor quickly senses equal and opposite charges in a strong electric field, generating an attraction force that pulls the drone towards the mesh surface. The advantage of the interactive polarity partition is that when the target drone comes into contact with the mesh surface, the polarity-opposite regions can form a transverse triboelectric potential gradient, which helps the drone to self-position itself to the center position on the mesh surface.

[0080] To avoid current surges during high-voltage boosting, this invention employs a multi-stage Martens generator series boosting method. The high-voltage module controls the boosting slope through parallel charging and series discharging. The controller dynamically adjusts the slope based on the real-time predicted value of remaining kinetic energy, ensuring both boosting speed and avoiding overload on the power supply system. After the electrostatic adsorption layer boosting is completed, the output voltage is maintained within a stable error range. The voltage feedback loop is adjusted in real time to suppress parasitic capacitance changes caused by contact with the body. The electromagnetic eddy current braking layer is located below the electrostatic layer, sharing the same flexible polyimide substrate. Its conductive path is a spiral multi-turn flat copper wire with consistent wire width and spacing. The wiring method ensures a closed loop exists in any cutting direction. The coil is driven by an AC current from a full-bridge inverter with a carrier frequency of 14 kHz. The AC magnetic field generates a changing magnetic flux on the conductive rotor blades or motor housing, forming a circulating current within the metal according to Faraday's law. The induced magnetic field generated by this circulating current is opposite in direction to the original magnetic field, thus producing a reverse braking torque on the rotating components. This braking torque can be approximated as... ,in, For braking torque, The rotor angular velocity, The eddy current braking coefficient depends on the number of coil turns, the amplitude of the driving current, the resistivity of the metal, and the magnetic field distribution. By measuring the coil port current and phase difference in real time, the controller calculates the instantaneous power and adjusts the driving current in a closed loop, so that the braking torque changes dynamically with the residual angular velocity, preventing excessive braking force from causing the machine body to shake.

[0081] The two-layer collaborative working mechanism is as follows: After the capture trigger signal arrives, the high-voltage power supply and inverter start simultaneously. The electrostatic layer quickly generates longitudinal traction, pulling the aircraft towards the center of the net. During the traction process, the vortex layer continuously applies reverse braking torque to reduce the rotor kinetic energy. Since the electrostatic traction force is inversely proportional to the square of the distance between the aircraft and the net, while the electromagnetic braking torque is proportional to the angular velocity, the braking layer contributes mainly to deceleration when the angular velocity is large in the early stage of capture. When the aircraft approaches the net, the traction force of the adsorption layer increases to complete the lock-in, achieving optimal torque matching. To quantify the energy decay effect of the adsorption layer and the braking layer, residual kinetic energy can be defined. ,in, Indicates the quality of the drone. Indicates translational velocity. The equivalent moment of inertia of the rotor is represented by online measurement. and The controller assesses the energy decay rate and decides when to reduce the voltage and drive current to reduce ineffective power consumption. Experimental results show that when the residual kinetic energy drops by more than 90%, continuing to maintain the high voltage will result in significant energy consumption and limited contribution to the locking effect. Therefore, the system enters the holding mode during this period, maintaining only the minimum holding voltage of the electrostatic layer and the pulse holding current of the braking layer.

[0082] In terms of material selection, the conductive fiber mesh uses graphene-coated aramid yarn, which has high flexibility and excellent conductivity; the polyimide substrate has a heat resistance temperature of over 450 degrees Celsius, ensuring that the inverter's long-term high-current drive does not cause thermal damage to the substrate; the copper wire uses a rectangular cross-section to improve the fill factor and reduce the skin effect; and the outer polyimide varnish layer ensures insulation.

[0083] The effectiveness evaluation used two types of UAVs: quadcopter and fixed-wing. In the quadcopter experiment, the initial velocity was 3.2 m / s, the angular velocity was 220 radians / s, the electrostatic layer voltage was increased to 45 kV, the eddy current driving current was 30 amperes peak, and the total power was 260 watts. Within 2 seconds, the rotor angular velocity dropped to 70 radians / s and the translational velocity dropped to 0.12 m / s, with the residual kinetic energy reduced by 92%, resulting in successful capture. In the fixed-wing experiment, the risk of vertical collision with the fixed-wing aircraft was lower due to electrostatic adsorption and traction, but the long wingspan structure increased the magnetic field coupling area. The test results showed that an increase of 15% in the driving current was sufficient to complete energy decay within 2.5 seconds, indicating that the solution has good adaptability to different aircraft types. Compared with traditional mechanical net solutions, this invention has three advantages: First, the electrostatic layer can instantly generate a volume force field, eliminating the need for aiming the projectile at the aircraft and providing a wider capture window; second, eddy current braking can rapidly decelerate the aircraft without contacting the rotor blades, avoiding damage to the aircraft structure; third, the flexible substrate is much lighter than a mechanical structure of the same size, minimizing its impact on the load of the accompaniment device. After long-term cyclic testing, the electrostatic layer has undergone more than 10,000 voltage boosts without breakdown, and the coil temperature rise remains below 60 degrees Celsius under a peak 2 kA current surge in the eddy current layer, meeting the reliability requirements for actual combat.

[0084] S5. Obtain the netting attitude data, compare it with the embedded vector to generate an update signal, and adjust the embedded vector and joint optimization algorithm.

[0085] After capture, the system enters a self-evolutionary phase, using the physical net-landing results to reverse-correct previous environmental perceptions and weight configurations, ensuring faster convergence to the effective navigation field in the next mission. Net-landing attitude data comes from a contact array beneath the net surface of the capture device. The contact array has a resolution of 64×64 and a sampling frequency of 1kHz. Each contact in the array has a built-in piezoresistive element; changes in piezoresistive resistance are converted from analog to digital to obtain a binary contact matrix. The controller first performs connected component analysis on the contact matrix to remove isolated noise points, and then calculates the contact centroid. With contact centroid principal axis direction This constitutes a five-dimensional attitude vector: ; in, The height is fixed at the mesh surface. To maintain dimensionality consistency with the embedded vectors, the system uses a two-layer perceptron to map the five-dimensional pose vectors to a 156-dimensional pose code. .

[0086] The set of embedded vectors is denoted as The contrast ratio is calculated using cosine similarity. ; In the formula, the dot product symbol represents the vector dot product. This represents the Euclidean norm. The closer it is to 1, the better. The higher the correlation between the node's behavior and its actual landing posture during the deception process, the better. The system will then... Sort the nodes and select the top 10% as high-contribution nodes. Then, apply weight amplification to these high-contribution nodes by multiplying their weights by a coefficient. For the remaining nodes, perform weight decay by multiplying by a coefficient. The amplification and attenuation factors satisfy: ; To ensure the conservation of total weight, among which This represents the updated weights. This operation is equivalent to providing the joint optimization algorithm with a new set of prior distributions, making it easier for the next annealing search to find low-energy candidate solutions near high-contribution nodes.

[0087] At the same time, the system will encode the attitude. Embedding the average vector with features from high-contribution nodes Calculate the difference vector This serves as the external gradient signal for the reinforcement learning policy network. At the start of the next training round, this is... Add a loss function: ; in To optimize the original loss term for the near-end strategy, This is a trade-off factor. Doing so explicitly encourages the policy network to generate a navigation field consistent with the actual landing attitude for orientation adjustments, thus shortening the policy convergence time.

[0088] To prevent overfitting due to noise from a single task, this invention writes the update signal into a distributed ledger. The ledger block header stores the timestamp and task number, while the block body stores the pose code, node weight correction coefficients, and gradient signal. The system reads the 16 most recent data blocks during the next task initialization and calculates a weighted average of the weight correction coefficients within them. This uses historical information to smooth the current correction, ensuring both the speed of self-evolution and offsetting the impact of abnormal posture samples. The weighted average weight decays exponentially over time, with the weight coefficient set to an exponential moving average of 0.8.

[0089] The performance evaluation results show that after five rounds of self-evolution, the initial value of the system's energy function decreased by about 40% compared to the first mission, the initial potential gradient alignment time was further reduced from 110ms to 75ms, the guidance time was shortened from 19s to 14s, and the peak power was reduced by about 8%. This indicates that attitude feedback effectively cleaned up low-contribution nodes and highlighted high-contribution paths, providing a better topology for navigation error injection.

[0090] Example: Five consecutive captures were performed, with pose encoding and weight correction recorded after each capture. At the start of the second task, 16 ledger records (including offline training data) were loaded, and the number of annealing search samples was reduced from 150 to 90 to achieve the same energy reduction. The fifth task required only 60 annealing search samples, and weight fine-tuning was completed in 3 rounds of reinforcement learning iterations, shortening the overall control cycle by nearly 30%. Experiments verified that the closed-loop design of pose feedback—ledger solidification—exponential smoothing has long-term stability and rapid adaptability.

[0091] Preferably, the netting attitude data is encoded into an attitude vector by a neural network, and the cosine similarity between the vector and the embedded vector is calculated to generate an update signal.

[0092] After acquisition, the system obtains the raw signals from the contact matrix, accelerometer, and microwave matrix instrument. These signals are first normalized by synchronization timestamps and then input into the "attitude encoding network." This network consists of one convolutional layer, two fully connected layers, and one normalization layer, with a total inference latency of approximately 3 milliseconds. The convolutional layer extracts local features from the contact matrix, outputting a tensor containing the local pressure distribution; the first fully connected layer concatenates the instantaneous data from the accelerometer, fusing the linear acceleration information of the attitude with the pressure quantization information; the second fully connected layer further compresses the dimension, fixing the output as a 156-dimensional real vector, called the attitude vector. .

[0093] The physical meaning is the comprehensive encoding of the machine's three-dimensional position, pitch angle, roll angle, and load distribution after landing on the network. The network is ultimately used... Normalization layer will Mapped to a unit hypersphere, ensuring length This normalization operation is to avoid the interference of amplitude factors when using cosine similarity in subsequent comparisons, and to retain only the direction information.

[0094] Embedded vector set Each one is pre-generated by the graph convolutional network during the deception phase. The same normalization process was applied. To quantify the correlation between landing pose and the contribution of each node, this invention calculates cosine similarity: ; in" " denotes the vector dot product; For the first Embed a navigation field node; This is guaranteed by normalization. Since the magnitudes of both vectors are both 1, the numerical range of the above equation is... . A value close to 1 indicates that the node contributes significantly to the actual captured attitude, while a value close to -1 indicates that the node's directional contribution is opposite during the deception process. The system then proceeds according to... Sort the nodes by value from largest to smallest, and select the top 10% as the set of high-contribution nodes. An update signal is then generated. : ; in Size of the set. Vector The direction represents the offset between the pose encoding vector and the mean of high-contribution nodes, and the magnitude is... This reflects the degree of difference between the two. If Approximately zero, proving that node embeddings can adequately characterize the captured pose; if A large value indicates that the node features do not adequately characterize the pose and need adjustment. Navigation field node weight correction is based on... Based on this, for high-contribution nodes, the weight is multiplied by an amplification factor: ; in This is an empirical coefficient, typically taken as 0.3. For low-contribution nodes, the weight is multiplied by a decay coefficient: ; After amplification and attenuation, the system renormalizes the weights to maintain a total weight sum of 1. This design can significantly strengthen high-contribution nodes while suppressing low-contribution nodes when the update signal differences are large; it also keeps the weights stable when the differences are small, preventing over-adjustment. It is worth noting that... Vectors not only affect weights but also feed back into the policy network of the joint optimization algorithm. A regularization term is added to the policy loss function: ; in This is a regularization weight. This encourages new strategies to generate navigation field modifications that are more consistent with the final capture pose, thus leading to faster convergence in the next round of deception.

[0095] The updated signal and corrected weights are written to the distributed ledger, and three fields are recorded in the block: attitude code. Weighting gain coefficient Table and difference vector The ledger block header includes a timestamp and task number, ensuring that subsequent tasks can quickly retrieve the most recent records. Upon the next startup, the system first reads the 16 most recent records, then performs a time-decay-weighted average of the weight gain coefficients to generate initial weights. The averaged attitude encoding mean is then injected into the policy network as a warm-up vector, making the network's initial policy closer to historically successful attitude captures.

[0096] In terms of effectiveness, the attitude feedback mechanism offers three advantages: First, by reviewing the navigation field model through physical capture results, misleading nodes can be eliminated in complex dynamic environments; second, injecting guidance vectors into the joint optimization algorithm significantly reduces the search space, with experiments showing a halving of the policy convergence steps; third, ledger solidification and time decay smoothing prevent a few anomalous samples from causing deviations in the global model. Experimental data from five rounds of tasks demonstrate that the initial value of the energy function decreases continuously, with each decrease being approximately 9%, and the initial gradient alignment time decreases from 110 milliseconds initially to 70 milliseconds in the fifth round, shortening the total guidance time to 14 seconds.

[0097] Example: In field testing, the capture mesh under the contact array had a resolution of 64×64 and a sampling frequency of 1kHz; the attitude coding network inference latency was 3 milliseconds. After the first round of tasks... , , Fifth round of tasks It has dropped to 0.11. , The weight distribution gradually stabilizes. The average inference latency of the policy network remains below 2 milliseconds, meeting the requirements for real-time control.

[0098] Preferably, the update signal is written to the distributed ledger after capture is completed, and loaded as the embedding vector and initial weights of the joint optimization algorithm when the next capture task is initialized.

[0099] In this invention, the distributed ledger is designed as a memory for a "capture-learn" closed loop, specifically recording the update signals, node weight correction coefficients, and policy regularization information obtained at the end of each task, ensuring that this information is traceable, immutable, and readable with low latency in a multi-node deployment environment. The ledger employs a lightweight Byzantine fault-tolerant consensus algorithm, generating a new block every 250 milliseconds. The block header includes the block height, the hash of the previous block, a timestamp, and a threshold signature, while the block body stores three core fields: The attitude encoding field records the captured attitude vector. (Length 156, normalized); The weight gain field contains the amplification factor. With attenuation coefficient Node index-value pairs; The regular gradient field records the difference vector. (Similarly normalized) and the regularization factor in this round .

[0100] To compress bandwidth, all three fields are stored in fixed-point format, with each element being 12 bits; the entire block size is approximately 3KB, which will not significantly burden local area network transmission. After consensus is reached on a new block, it is broadcast to all navigation field nodes and accompanying devices, persisted locally to solid-state storage, and immediately enters a searchable state.

[0101] During the next capture task initialization, the system retrieves the 16 most recent records through the ledger query interface. To balance historical information with the latest scene features, the system uses an exponential moving average to generate the Warm-start value for the navigation field node weights. Specifically, the records are traversed in reverse chronological order. The amplification-decrease table in each record is first reflected in the current weight vector. Then according to the coefficient Superimposed, The time decay factor is set to 0.8. The mathematical expression is: ; in Representing the The corrected node weight vector of each record. The initial weights for this task are set. Since the weights have already been normalized, This also satisfies the condition that the sum is 1. This preserves the advantage of high-contribution nodes in recent tasks while avoiding excessive bias towards isolated outlier samples.

[0102] The policy network Warm-start uses the attitude encoding mean. mean of regular vectors Both are smoothed using an exponential decay coefficient, resulting in... ; Used as the bias initialization term for the weights of the first layer of the policy network. Used to set the initial gradient direction for the regularization loss. Practical experience shows that Warm-start can reduce the average number of convergence iterations of the policy network by 45% and the number of annealing search samplings by 40%, significantly improving real-time performance.

[0103] In terms of security, to prevent malicious nodes from uploading forged update signals, the ledger uses threshold signatures; a block is only accepted by the network when more than two-thirds of the navigation field nodes jointly sign it. Each block verification simultaneously calculates the Hamming distance between the newly added weight table and the previous weight table. If the distance exceeds a set threshold, manual review is triggered to prevent extreme abnormal data from corrupting the global model. Ledger content transmission uses session key encryption, with the key generated by the previous round of consensus protocol negotiation, ensuring that update signals cannot be tampered with even if communication is eavesdropped on.

[0104] In terms of performance, five rounds of self-evolution experiments showed that after using Warm-start, the initial value of the energy function was reduced by about 40% compared to the first task, the initial potential gradient alignment time was reduced from 110 milliseconds to 70 milliseconds, the total guidance time was shortened from 19 seconds to 14 seconds, and the peak power was reduced by another 8%. In the multi-target demonstration, the system maintained its own ledger branch for each of the three drones. Warm-start initialization ensured that the three decoy trajectories did not interfere with each other, and finally completed the capture sequentially within 45 seconds, verifying the parallel scalability of the partitioned ledger and time decay strategy.

[0105] like Figure 2 As shown, a joint optimization system for UAV navigation error implantation and target acquisition is used to implement the aforementioned joint optimization method for UAV navigation error implantation and target acquisition. The system includes: The environmental acquisition module is used to collect optical data and wireless signal data and generate a unified time base through a distributed consensus mechanism. The environmental acquisition module includes: a dual-band camera + software radio front-end; a rubidium clock-FPGA time system; and fiber optic interconnection voting to generate a unified timestamp.

[0106] The model generation module is used to construct a spatial environment model under a unified time reference to obtain embedding vectors and initial parameters of the navigation field. The model generation module includes: edge GPUs to voxelize synchronous point clouds and radio frequency features, and graph convolution inference to derive node embeddings and initial weights.

[0107] The illusion transmission module is used to synchronously transmit illusion signals based on the embedded vector, the real-time status of the target UAV, and the initial parameters of the navigation field, and to implant navigation errors in the airspeed channel, visual channel, and magnetic compass channel of the target UAV. The illusion transmission module includes an ultrasonic array, a digital micromirror array, and an orthogonal coil sharing a trigger pulse bus. The hardware phase-locked loop jitter is less than 10 microseconds, and it synchronously outputs three types of illusion signals.

[0108] The optimization and update module is used to couple the predicted state of the target UAV with the navigation error using a joint optimization algorithm to update the navigation field parameters, and to continuously drive the illusion emission module to emit illusion signals to guide the target UAV to the predetermined capture position using the updated navigation field parameters. The optimization and update module includes: CPU annealing search to give candidate weights; GPU reinforcement learning to fine-tune the weights and rewrite the emission register in real time.

[0109] The capture execution module is used to determine, based on navigation field parameters and the real-time status of the target UAV, that the target UAV is located in a predetermined space and its flight speed does not exceed a predetermined speed threshold. When this determination is made, the capture device, which includes an electrostatic adsorption layer and an electromagnetic eddy current braking layer, is activated to perform deceleration capture and outputs net-landing attitude data. The capture execution module includes: a high-voltage interactive polarity mesh surface adsorption body; a flexible spiral coil alternating magnetic field applying eddy current braking to the rotor; and a contact array for acquiring net-landing attitude.

[0110] The feedback update module compares the netting pose data with the embedding vectors to generate an update signal, and adjusts the embedding vectors and joint optimization algorithm parameters based on the update signal for subsequent capture. The feedback update module includes: calculating the similarity between pose encoding and node embedding to generate the update signal; writing the update signal to the distributed ledger and loading it as a warm-start weight on the next power-on.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

[0112] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A joint optimization method for UAV navigation error implantation and target acquisition, characterized in that, Includes the following steps: Collect UAV data and unify its time reference; obtain the embedding vector and initial parameters of the navigation field through the spatial environment model. Based on the initial parameters and embedding vector of the navigation field, combined with the real-time status of the target UAV, an illusion signal is synchronously emitted to implant navigation error into the target UAV; The navigation field parameters are updated using a joint optimization algorithm, and illusion signals are continuously emitted to guide the target UAV to the predetermined capture position. Based on the navigation field parameters and the real-time status of the target UAV, determine whether to activate the capture device to perform deceleration capture; Acquire netting attitude data, compare with the embedded vector to generate update signals, and adjust the embedded vector and joint optimization algorithm.

2. The joint optimization method for UAV navigation error implantation and target acquisition according to claim 1, characterized in that, The process of constructing a space environment model includes: A three-dimensional coordinate system is generated by reconstructing three-dimensional point clouds, meshing, and registering coordinates from optical data and wireless signal data. An embedding vector is then established within the three-dimensional coordinate system using the navigation field node positions as indices.

3. The joint optimization method for UAV navigation error implantation and target acquisition according to claim 1, characterized in that, The illusion signals include: Acoustic illusion signal, wherein the acoustic illusion signal is emitted by a phased-array ultrasonic directional sound pressure field; Optical illusion signal, wherein the optical illusion signal is generated by projecting environmental model feature patterns from a digital micromirror array; A magnetic field illusion signal, wherein the magnetic field illusion signal is generated by an alternating magnetic field by a dual-coil device.

4. The joint optimization method for UAV navigation error implantation and target acquisition according to claim 1, characterized in that: The timing of the transmission of the illusion signal is synchronized with a unified time reference to ensure that the navigation error remains consistent in time across the target UAV's airspeed channel, visual channel, and magnetic compass channel.

5. The joint optimization method for UAV navigation error implantation and target acquisition according to claim 1, characterized in that, The joint optimization algorithm includes: Annealing search, which generates a set of candidate navigation field parameters; Reinforcement learning, wherein the reinforcement learning continuously adjusts the weights of the candidate navigation field parameter set based on the real-time residual between the predicted state information of the target UAV and the navigation error information; The annealing search and reinforcement learning are performed sequentially.

6. The joint optimization method for UAV navigation error implantation and target acquisition according to claim 1, characterized in that, The activation of the capture device is based on the following determination: The spatial error of the navigation field parameter output is no greater than the preset spatial threshold. The speed error of the navigation field parameter output is not greater than the preset speed threshold; If both the space threshold and the speed threshold are met, a capture trigger signal is issued to start the capture device to perform deceleration capture.

7. The joint optimization method for UAV navigation error implantation and target acquisition according to claim 1, characterized in that, The capture device includes: An electrostatic adsorption layer, wherein the electrostatic adsorption layer is composed of an alternating polarity partitioned conductive fiber mesh. An electromagnetic eddy current braking layer is provided, wherein multiple turns of wire are embedded in a flexible substrate and eddy currents are generated by AC excitation to apply a reverse braking torque to the rotor of a target UAV.

8. The joint optimization method for UAV navigation error implantation and target acquisition according to claim 1, characterized in that: The net-landing attitude data is encoded into an attitude vector by a neural network, and then cosine similarity calculation is performed between the vector and the embedded vector to generate an update signal.

9. The joint optimization method for UAV navigation error implantation and target acquisition according to claim 8, characterized in that: The update signal is written into the distributed ledger after capture is completed, and is loaded as the initial weights of the embedding vector and joint optimization algorithm during the initialization of the next capture task.

10. A joint optimization system for UAV navigation error implantation and target acquisition, applied to the method described in any one of claims 1 to 9, characterized in that, include: The environmental acquisition module is used to collect optical data and wireless signal data and generate a unified time reference through a distributed consensus mechanism. The model generation module is used to construct a spatial environment model under a unified time reference to obtain the embedding vector and initial parameters of the navigation field; The illusion emission module is used to synchronously emit illusion signals based on the embedded vector, the real-time status of the target UAV, and the initial parameters of the navigation field, thereby implanting navigation errors in the airspeed channel, visual channel, and magnetic compass channel of the target UAV. The optimization and update module is used to use a joint optimization algorithm to couple the predicted state of the target UAV with the navigation error to update the navigation field parameters, and use the updated navigation field parameters to continuously drive the illusion emission module to emit illusion signals to guide the target UAV to the predetermined capture position. The capture execution module is used to determine, based on the navigation field parameters and the real-time status of the target UAV, that the target UAV is located in a predetermined space and its flight speed does not exceed a predetermined speed threshold. When this determination is made, the capture device, which includes an electrostatic adsorption layer and an electromagnetic eddy current braking layer, is activated to perform deceleration capture and outputs the net-landing attitude data. The feedback update module is used to compare the net-landing attitude data with the embedding vector to generate an update signal, and adjust the embedding vector and joint optimization algorithm parameters according to the update signal for subsequent capture.

Images