Unmanned aerial vehicle soft landing control method based on flight path prediction and multi-source cooperative decoy

Abstract

The invention discloses an unmanned aerial vehicle soft landing control method based on track prediction and multi-source cooperative decoy, and the method comprises the steps: introducing a communication protocol recognition mechanism in a blocking process of an instruction link of an unmanned aerial vehicle, and achieving the conversion from extensive broadband suppression to protocol precise interference; the problems that in the prior art, power consumption is high, interference on legal communication is large, and the frequency hopping unmanned aerial vehicle resisting effect is insufficient can be effectively solved. In the scheme, signal integrity is guaranteed through monitoring and preprocessing, an unmanned aerial vehicle link protocol is identified through feature extraction and protocol library matching, and then a targeted interference template is generated, so that the interference effect can accurately fall on key positions such as a leader sequence, a CRC field, a subcarrier or a control frame and the like; therefore, synchronization, verification or data transmission can be destroyed under the condition of low power, and accidental damage of broadband noise to services such as co-frequency Wi-Fi and Bluetooth is avoided.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) soft landing control technology, and in particular to a UAV soft landing control method based on trajectory prediction and multi-source cooperative deception. Background Technology

[0002] With the widespread application of unmanned aerial vehicles (UAVs) in military reconnaissance, border patrol, urban security, aerial surveying, and logistics, their numbers and complexity continue to increase, posing new challenges to airspace security and ground public safety. When UAVs experience communication link loss, flight control system malfunctions, or enter controlled defense zones, effectively controlling them and ensuring their safe landing has become a crucial research direction in UAV traffic management and counter-UAV technology. Compared to traditional methods of direct shooting or violent jamming, achieving soft landing control without damage not only helps reduce ground losses and safety hazards but also provides conditions for subsequent evidence collection, analysis, and reuse, thus possessing significant engineering and application value.

[0003] Existing technologies primarily control drones through single-signal jamming or deception. For example, GPS / BeiDou navigation spoofing technology alters the drone's trajectory by forging positioning signals; communication link suppression and remote control interference force the drone into return-to-home or hovering mode by blocking its connection with the ground station; some research attempts to influence its attitude using visual interference or magnetic field disturbances. However, these methods have significant shortcomings: first, single-signal spoofing is easily detected and resisted by drones with multi-source fusion navigation capabilities; second, brute-force interference often leads to drone loss of control and crashes, increasing safety risks; and third, existing trajectory control methods lack prediction of drone flight dynamics and trajectory trends, easily triggering abrupt maneuvers that may induce anomaly detection or self-destruct mechanisms, resulting in failed soft landings.

[0004] Therefore, in the controlled soft landing process of UAVs, insufficient trajectory prediction accuracy, poor coordination consistency of multi-source deception, and weak smoothness of dynamic guidance have become problems that urgently need to be solved. Summary of the Invention

[0005] This application provides a method for soft landing control of unmanned aerial vehicles (UAVs) based on trajectory prediction and multi-source cooperative deception, aiming to solve the problems of insufficient trajectory prediction accuracy, poor consistency of multi-source deception, and weak smoothness of dynamic guidance during the controlled soft landing process of UAVs.

[0006] Firstly, a soft landing control method for unmanned aerial vehicles (UAVs) based on trajectory prediction and multi-source cooperative deception, the method comprising:

[0007] Acquire multimodal state information of the target UAV and identify the model and flight control logic. The multimodal state information includes position, velocity, attitude, trajectory sequence, and navigation and anomaly detection features.

[0008] The state information is used to perform hierarchical trajectory prediction, which includes short-term prediction based on a six-degree-of-freedom dynamic model, medium- and long-term prediction based on a temporal deep network, and game-theoretic prediction based on the adversarial maneuver hypothesis. The different prediction results are then fused to generate a decoy target corridor and descent profile.

[0009] Based on the prediction results, a multi-source collaborative virtual signal is generated, which includes at least a satellite navigation virtual signal, a barometric altitude virtual quantity, an inertial measurement unit perturbation quantity, and a visual landmark reference.

[0010] The multi-source collaborative virtual signal is subjected to consistency verification and constraint shaping to meet the requirements of spatiotemporal alignment, dynamic reachability and sensor noise statistical characteristics.

[0011] The virtual signals are dynamically weighted according to the flight stage of the UAV, and a smooth transition is maintained when the weights change.

[0012] A gradual perturbation is applied to the virtual signal to guide the drone to gradually deviate from its original trajectory and enter the decoy target corridor;

[0013] Closed-loop adaptive adjustment of prediction parameters, signal generation parameters, and weight vectors is performed based on the real-time feedback status of the UAV.

[0014] When the drone enters the preset corridor terminal, it provides its flight control system with landing trigger conditions and landing reference information, enabling the drone to complete a soft landing.

[0015] Optionally, in the above scheme, the fusion method of the hierarchical trajectory prediction is as follows:

[0016] By using dynamic prediction as the process model input for the extended Kalman filter, the prediction mean and covariance are established.

[0017] The time series prediction results are used as the observation prediction input, and the residuals are compared with the dynamic prediction. When the residuals are lower than the preset threshold, an update is performed, and when they are higher than the threshold, a fault-tolerant update strategy is adopted.

[0018] The game prediction results are used as a constraint set or penalty cost term to guide the final prediction trajectory;

[0019] The covariance cross-update and Bayesian weight allocation methods are used to jointly optimize different prediction results.

[0020] Optionally, in the above scheme, the game-theoretic prediction includes:

[0021] The maneuvering strategy of the target UAV is modeled as a finite set of actions, which includes at least hovering, acceleration to break through defenses, deceleration to hover, climbing to escape, and return to home.

[0022] Establish a minimax game model so that the deception system takes minimizing trajectory deviation as its payoff function, while the drone takes increasing trajectory deviation or successfully returning to base as its payoff function.

[0023] The Stackelberg leader-follower game model is adopted, with the decoy system as the leader and the drone as the follower. The set of adversarial trajectories is generated by solving the game equilibrium.

[0024] The set of adversarial trajectories is subjected to a reachability test, which includes state space boundary, acceleration upper limit and energy margin constraint;

[0025] The trajectory envelope that passes the test is cross-compared with other prediction results to obtain the corrected game prediction sequence.

[0026] Optionally, in the above scheme, the consistency verification and constraint shaping includes:

[0027] Error statistical models are established for GNSS, IMU, barometric and visual signals, respectively. The models include Gaussian noise, low-frequency bias and sudden interference.

[0028] Calculate the joint residual vector between different signal sources and compare it with a preset statistical threshold;

[0029] The residual results are input into the constraint optimizer, and a constraint solution is generated using quadratic programming or convex optimization methods.

[0030] The amplitude and timing of the virtual signal are adjusted according to the constraint solution to ensure that the multi-source signals maintain consistency in the statistical sense of covariance.

[0031] Optionally, in the above scheme, the dynamic weight coordination and phased guidance include:

[0032] The flight process of the drone is divided into four stages: long-range high altitude, medium-altitude transition, low-altitude approach, and terminal landing.

[0033] During the long-range, high-altitude phase, the weight of GNSS signals shall not be less than 70%;

[0034] During the hollow transition phase, the weight of the air pressure signal shall be no less than 20%, and the weight of the IMU signal shall be no less than 25%.

[0035] During the low-altitude approach phase, the weight of visual signals shall be no less than 40%, and the weight of barometric signals shall be no less than 30%.

[0036] During the terminal landing phase, the policy network based on the Markov decision process adjusts the weights of various signals in real time.

[0037] The transition between each stage is managed by a finite state machine to achieve a smooth transition.

[0038] Optionally, in the above scheme, the progressive perturbation control satisfies the following constraints:

[0039] The amplitude of a single-step perturbation, after statistical testing, does not exceed twice the standard deviation of the sensor noise.

[0040] The first-order difference and the second-order difference at adjacent time points are both less than the preset threshold;

[0041] The pseudorange, carrier phase, and code phase of GNSS signals are independently constrained in both the frequency and time domains;

[0042] The bias, scaling factor, and random walk components of the IMU signal are limited by amplitude and bandwidth, respectively.

[0043] Visual signal perturbations are generated through pixel affine transformation and illumination perturbation, and are constrained by the upper limit of projection error.

[0044] Optionally, in the above scheme, the closed-loop adaptive adjustment includes:

[0045] Construct a feature vector based on the difference between the observed trajectory and the predicted trajectory;

[0046] The feature vector is input into a reinforcement learning agent, which employs a deep Q-network or a near-end policy optimization algorithm.

[0047] Based on the agent output, the parameters of the prediction module, signal generation module, and weight allocation module are updated in real time.

[0048] During the update process, a sliding window memory buffer and a priority experience replay mechanism are used.

[0049] When the drone exhibits abnormal maneuvers, the parameters are rolled back to the previous stable solution.

[0050] Optionally, in the above scheme, the near-ground landing triggering and execution includes:

[0051] Establish a geometric model of the safe landing zone in the target area, the model including open areas, runways, or artificial landmark areas;

[0052] Landing trigger conditions are set based on the drone's altitude and speed thresholds, and the conditions include return-to-home, low battery, or mission completion logic.

[0053] After the triggering conditions are met, glide path parameters are generated, including heading alignment angle, descent rate curve and contact point coordinates;

[0054] During landing, adjust the drone's descent rate, throttle, and attitude input to conform to the glide path parameters.

[0055] Virtual visual markers or optical landmarks are set up in the contact point area to assist the drone in completing the grounding determination.

[0056] Optionally, in the above scheme, the generation and transmission of the multi-source collaborative virtual signal includes:

[0057] The GNSS channel generates code phase, carrier phase, and Doppler frequency shift, and performs bandwidth shaping and power spectrum constraint.

[0058] A virtual atmospheric model of pressure channel variation with altitude is constructed, which includes temperature gradient, humidity correction and wind speed disturbance;

[0059] The IMU channel injects slowly varying bias, low-frequency random walk, and band-limited high-frequency noise into the triaxial gyroscope and accelerometer.

[0060] The visual channel generates a set of virtual optical landmarks and performs projection transformations based on camera parameters to obtain continuous virtual image frames;

[0061] The multi-channel signals are synchronized with a unified clock, including GPS clock alignment, hardware clock correction and delay compensation;

[0062] The alignment error of the multi-channel transmission is no greater than 1 microsecond and is compensated in real time through a phase-locked loop.

[0063] In a second aspect, a computer program product includes a computer program / instructions that, when executed by a processor, perform the following steps:

[0064] Acquire multimodal state information of the target UAV and identify the model and flight control logic. The multimodal state information includes position, velocity, attitude, trajectory sequence, and navigation and anomaly detection features.

[0065] The state information is used to perform hierarchical trajectory prediction, which includes short-term prediction based on a six-degree-of-freedom dynamic model, medium- and long-term prediction based on a temporal deep network, and game-theoretic prediction based on the adversarial maneuver hypothesis. The different prediction results are then fused to generate a decoy target corridor and descent profile.

[0066] Based on the prediction results, a multi-source collaborative virtual signal is generated, which includes at least a satellite navigation virtual signal, a barometric altitude virtual quantity, an inertial measurement unit perturbation quantity, and a visual landmark reference.

[0067] The multi-source collaborative virtual signal is subjected to consistency verification and constraint shaping to meet the requirements of spatiotemporal alignment, dynamic reachability and sensor noise statistical characteristics.

[0068] The virtual signals are dynamically weighted according to the flight stage of the UAV, and a smooth transition is maintained when the weights change.

[0069] A gradual perturbation is applied to the virtual signal to guide the drone to gradually deviate from its original trajectory and enter the decoy target corridor;

[0070] Closed-loop adaptive adjustment of prediction parameters, signal generation parameters, and weight vectors is performed based on the real-time feedback status of the UAV.

[0071] When the drone enters the preset corridor terminal, it provides its flight control system with landing trigger conditions and landing reference information, enabling the drone to complete a soft landing.

[0072] Compared with the prior art, this application has at least the following beneficial effects:

[0073] Based on further analysis and research into existing technical problems, this application recognizes the issues of insufficient trajectory prediction accuracy, poor coordination consistency among multi-source decoys, and weak smoothness of dynamic guidance during controlled soft landing of UAVs. By introducing a communication protocol identification mechanism during UAV command link blocking, it achieves a shift from "extensive broadband suppression" to "precise protocol interference," effectively addressing the problems of high power consumption, significant interference with legitimate communications, and insufficient effectiveness against frequency-hopping UAVs in existing technologies. In this scheme, signal integrity is first ensured through monitoring and preprocessing. Then, the UAV link protocol is identified through feature extraction and protocol library matching, generating targeted interference templates. This ensures that the interference accurately targets key positions such as preamble sequences, CRC fields, subcarriers, or control frames, thereby disrupting synchronization, verification, or data transmission even under low-power conditions, avoiding accidental damage to services like Wi-Fi and Bluetooth on the same frequency due to broadband noise. Simultaneously, through a frequency-hopping prediction model and scheduling mechanism, the interference signal can follow the UAV's frequency hopping in real time, achieving continuous blocking and compensating for the ineffectiveness of traditional fixed-frequency interference against frequency-hopping anti-jamming UAVs. Furthermore, the closed-loop control mechanism dynamically optimizes the interference waveform, power, and time slot allocation based on real-time link state variables, ensuring the interference remains stable and reliable in complex electromagnetic environments. Therefore, this invention significantly reduces collateral effects and energy consumption while guaranteeing interference effectiveness, comprehensively solving the problems of low interference efficiency, significant collateral damage, and insufficient real-time performance in the prior art. Attached Figure Description

[0074] Figure 1 This is a flowchart illustrating a method for soft landing control of unmanned aerial vehicles based on trajectory prediction and multi-source cooperative deception, provided as an embodiment of this application. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0076] In the description of this application: unless otherwise stated, "a plurality of" means two or more. The terms "first," "second," "third," etc., in this application are intended to distinguish the objects referred to and do not have any special meaning in terms of technical connotation (e.g., they should not be construed as an emphasis on importance or order). Expressions such as "including," "comprising," and "having" also mean "not limited to" (certain units, components, materials, steps, etc.).

[0077] In one embodiment, such as Figure 1 As shown, a soft landing control method for unmanned aerial vehicles (UAVs) based on trajectory prediction and multi-source cooperative deception is provided, including the following steps:

[0078] Acquire multimodal state information of the target UAV and identify the model and flight control logic. The multimodal state information includes position, velocity, attitude, trajectory sequence, and navigation and anomaly detection features.

[0079] The state information is used to perform hierarchical trajectory prediction, which includes short-term prediction based on a six-degree-of-freedom dynamic model, medium- and long-term prediction based on a temporal deep network, and game-theoretic prediction based on the adversarial maneuver hypothesis. The different prediction results are then fused to generate a decoy target corridor and descent profile.

[0080] Based on the prediction results, a multi-source collaborative virtual signal is generated, which includes at least a satellite navigation virtual signal, a barometric altitude virtual quantity, an inertial measurement unit perturbation quantity, and a visual landmark reference.

[0081] The multi-source collaborative virtual signal is subjected to consistency verification and constraint shaping to meet the requirements of spatiotemporal alignment, dynamic reachability and sensor noise statistical characteristics.

[0082] The virtual signals are dynamically weighted according to the flight stage of the UAV, and a smooth transition is maintained when the weights change.

[0083] A gradual perturbation is applied to the virtual signal to guide the drone to gradually deviate from its original trajectory and enter the decoy target corridor;

[0084] Closed-loop adaptive adjustment of prediction parameters, signal generation parameters, and weight vectors is performed based on the real-time feedback status of the UAV.

[0085] When the drone enters the preset corridor terminal, it provides its flight control system with landing trigger conditions and landing reference information, enabling the drone to complete a soft landing.

[0086] This embodiment proposes a soft landing control method for unmanned aerial vehicles (UAVs) based on trajectory prediction and multi-source collaborative deception, covering the entire process from UAV state perception, trajectory prediction, virtual signal generation, consistency verification, dynamic guidance, progressive control, closed-loop adaptive adjustment to final soft landing triggering. This independent claim is the core of the entire patent solution, solving the problem in existing technologies where UAV crashes or deception failures are caused by single-signal deception, discontinuous interference, or lack of closed-loop control.

[0087] In terms of implementation, the method first acquires the target UAV's state information and identifies its flight control logic through multimodal perception. State information includes position, velocity, attitude, flight path sequence, and navigation and anomaly detection features. This information can be acquired through radar measurement, radio frequency detection, optical recognition, and infrared imaging, among others. Identifying the flight control logic requires pattern matching between an aircraft model database and a flight control behavior feature database. For example, by acquiring the frequency hopping characteristics of the UAV's remote control link through radio frequency monitoring and comparing it with the database, the possible flight control manufacturer and anomaly detection mechanism can be determined. This approach ensures the deception scheme is highly targeted, avoiding failure due to "blind deception."

[0088] After assessing the UAV's status, the system enters a hierarchical trajectory prediction phase. This prediction comprises three levels: short-term prediction based on a six-degree-of-freedom dynamics model to ensure the physical plausibility of the trajectory; medium- to long-term prediction based on a temporal deep network to capture the UAV's mission trends and long-term planning; and game-theoretic prediction based on the adversarial maneuver assumption to simulate the UAV's countermeasures when encountering interference. The three prediction results are fused to output a target corridor and descent profile with a confidence distribution. This multi-level prediction avoids the problem of excessive prediction bias from a single model, improving the robustness of trajectory deception.

[0089] Guided by the prediction results, the system generates multi-source collaborative virtual signals. The GNSS virtual signal, by falsifying data conforming to ephemeris and orbital patterns, causes the UAV to gradually deviate from its original trajectory at the positioning level; the barometric altitude virtual signal, by simulating pressure changes in the real atmospheric environment, causes the UAV's perceived altitude to naturally decrease; the IMU virtual signal, by superimposing low-amplitude offsets and random walks on three-axis gyroscope and accelerometer data, ensures consistency with the GNSS displacement trend; and the visual signal is generated through virtual landmark projection or image simulation, ensuring that the UAV's visual navigation module also receives a consistent reference. The technical effect of this step is to ensure that the virtual data from different sensor sources possess both physical and numerical realism and consistency.

[0090] To avoid conflicts between multi-source signals, the system employs a consistency check and constraint shaping mechanism. Specifically, it performs cross-source residual comparisons of GNSS, IMU, barometric pressure, and visual data, comparing the results to preset thresholds. If the residuals exceed normal ranges, the system automatically corrects the amplitude and timing of the virtual signal to ensure statistical consistency among the multi-source signals. For example, if the GNSS descent rate is too fast while the barometric pressure signal descent is slow, the system adjusts the descent gradient of the GNSS signal to maintain coordination between the two. The technical effect of this mechanism is to prevent the UAV flight control system from detecting anomalies during multi-source data fusion, thereby enhancing deception stealth.

[0091] During the drone guidance process, the system also incorporates a dynamic weighted coordination and phased guidance mechanism. Specifically, GNSS signals have the highest weight during the high-altitude phase; during the mid-altitude transition phase, the weight of barometric pressure and IMU signals gradually increases; and during the low-altitude approach phase, visual and barometric pressure signals dominate. The transitions between different phases are managed through a finite state machine, ensuring smooth weight changes and avoiding abrupt shifts. This phased guidance technique ensures that the virtual signals match the drone's navigation habits at different flight stages, making the deception process natural and difficult to detect.

[0092] Furthermore, the method employs progressive perturbation control, applying small-amplitude perturbations to the virtual signal. Both the perturbation amplitude and rate of change are constrained, with a single-step perturbation not exceeding twice the sensor noise, and the rate of change remaining continuous between adjacent time points. This control method smooths the UAV's flight trajectory deviation process, preventing the triggering of the flight control system's safety protection logic, thus achieving natural guidance.

[0093] During the drone's flight, the system also incorporates closed-loop adaptive adjustment. By collecting the drone's feedback status in real time and comparing it with the prediction results, the system updates the signal generation parameters and prediction weights using reinforcement learning agents or adaptive laws. When the drone deviates from its course or performs abnormal maneuvers, the system can quickly adjust its deception strategy, or even roll back to stable parameters, ensuring the deception remains effective. The technical effect of this step is to enhance adaptability to complex environments and different drone models, thereby improving the robustness of the deception.

[0094] As the drone approaches the pre-designated landing area, the system triggers soft landing control. This is achieved by injecting a virtual trigger signal when the drone meets specific conditions (such as return-to-home logic, low battery, or mission completion), causing it to automatically enter landing mode. The system simultaneously provides virtual glide path parameters and ground visual references to guide the drone in autonomously completing a soft landing. Unlike traditional forced landings, this method is executed by the drone's own flight control system, resulting in a smooth and stealthy landing process.

[0095] This method constructs a full-process, closed-loop control scheme for UAV soft landing through a series of steps, including multimodal state perception, hierarchical prediction, multi-source collaborative virtual signal generation, consistency shaping, dynamic weight allocation, gradual guidance, closed-loop adaptation, and natural landing triggering. Its technical effects are reflected in three aspects: First, it achieves highly consistent deception using multi-source signals, making it difficult for the UAV to detect; second, hierarchical prediction and gradual guidance ensure the naturalness and continuity of trajectory control; and third, closed-loop adjustment and soft landing triggering ensure a safe and stable landing for the UAV. This scheme effectively solves the problems of easy identification of single deception methods and the risk of crashes due to forced landings in existing technologies, demonstrating significant innovation and practical value.

[0096] In this embodiment, the fusion method of the hierarchical trajectory prediction is as follows:

[0097] By using dynamic prediction as the process model input for the extended Kalman filter, the prediction mean and covariance are established.

[0098] The time series prediction results are used as the observation prediction input, and the residuals are compared with the dynamic prediction. When the residuals are lower than the preset threshold, an update is performed, and when they are higher than the threshold, a fault-tolerant update strategy is adopted.

[0099] The game prediction results are used as a constraint set or penalty cost term to guide the final prediction trajectory;

[0100] The covariance cross-update and Bayesian weight allocation methods are used to jointly optimize different prediction results.

[0101] This embodiment further defines the fusion method of hierarchical trajectory prediction, ensuring that different prediction results can be integrated through a scientific mechanism to generate a final predicted trajectory with higher confidence and stronger physical plausibility. In the process of UAV deception and guidance, the accuracy of trajectory prediction directly determines whether the deception strategy is natural and credible. If the prediction deviation is too large, it is very easy to trigger the abnormal detection logic of the UAV flight control system. Therefore, this claim provides a detailed description of the prediction fusion method.

[0102] In its implementation, the system first uses dynamic prediction as the input to the process model of the extended Kalman filter. As a highly nonlinear system, the flight dynamics of a UAV are influenced by multiple factors, including aerodynamics, propulsion, and control laws. To ensure the accuracy of short-term predictions, a dynamic model based on six degrees of freedom equations is introduced. By linearizing the process model and noise term, the extended Kalman filter can provide accurate trajectory predictions within a short timeframe, outputting the state mean and covariance matrix. This result serves as the basis for subsequent time-series and game-theoretic predictions, ensuring that the trajectory is physically achievable.

[0103] Subsequently, the system uses the time-series prediction results based on deep neural networks as the input for observation prediction. The deep learning model, trained on the historical trajectories of the UAV, can extract its potential flight patterns, especially medium- to long-term mission trends. When the time-series predictions are compared with the dynamic predictions, a residual vector is generated. If the residual is within a preset threshold, it indicates good consistency between the two, and the dynamic prediction results can be corrected by updating the Kalman gain. If the residual exceeds the threshold, it indicates that there may be anomalies in the time-series prediction, such as prediction drift due to insufficient dataset. In this case, the system will reduce its weight or even temporarily discard the input. This residual gating mechanism makes the system fault-tolerant during the fusion process, preventing the bias of a single model from being amplified infinitely.

[0104] To further enhance the adaptability of trajectory prediction, this invention introduces game theory prediction results as constraints. Game theory prediction considers the behavioral strategies of the UAV in adversarial environments and can simulate its possible countermeasures. Such prediction results are typically input into the fusion model in the form of constraint sets or penalty cost terms to limit the deviation range of the predicted trajectory. For example, when game theory prediction indicates that the UAV may perform a return-to-home action, the fusion model increases the cost of deviating from the return-to-home trajectory, making the final prediction more consistent with the UAV's potential behavioral logic. This mechanism ensures that the fusion results remain reliable even in complex situations.

[0105] In the process of fusing multiple prediction results, the system employs a covariance cross-update and Bayesian weight allocation method. Specifically, the covariance cross-update achieves dynamic weight adjustment by exchanging uncertainty information among different prediction models. When dynamic predictions have low uncertainty, their results will have higher weights; while in long-term predictions, time-series models can provide trend information and therefore receive more weight. The Bayesian method treats the prediction results of different models as probability distributions and achieves optimal weighting through posterior probability calculation. This method enables the prediction results to adapt to environmental changes and improves the reliability of the final trajectory.

[0106] From a technical perspective, this implementation method ensures the prediction results in terms of physical rationality, statistical stability, and adversarial adaptability. Dynamic prediction ensures short-term trajectory accuracy, time-series prediction compensates for long-term trends, game-theoretic prediction provides constraints in adversarial scenarios, and covariance cross-update and Bayesian assignment ensure the reasonable fusion of results from different models. Overall, this claim addresses the shortcomings of single prediction methods, such as large biases, instability, and lack of adversarial adaptability, enabling the deception system to possess greater accuracy and stealth when controlling drones into descent corridors.

[0107] In this embodiment, the game-theoretic prediction includes:

[0108] The maneuvering strategy of the target UAV is modeled as a finite set of actions, which includes at least hovering, acceleration to break through defenses, deceleration to hover, climbing to escape, and return to home.

[0109] Establish a minimax game model so that the deception system takes minimizing trajectory deviation as its payoff function, while the drone takes increasing trajectory deviation or successfully returning to base as its payoff function.

[0110] The Stackelberg leader-follower game model is adopted, with the decoy system as the leader and the drone as the follower. The set of adversarial trajectories is generated by solving the game equilibrium.

[0111] The set of adversarial trajectories is subjected to a reachability test, which includes state space boundary, acceleration upper limit and energy margin constraint;

[0112] The trajectory envelope that passes the test is cross-compared with other prediction results to obtain the corrected game prediction sequence.

[0113] The core of this game-theoretic prediction method lies in modeling the interaction between the drone and the decoy system to predict the drone's possible maneuvers when interfered with or suspected of being abnormal. Existing technologies often neglect the drone's countermeasure logic, resulting in unrealistic predicted trajectories that are easily identified by the target system. This invention forms a complete game-theoretic prediction system through finite action set modeling, minimax game theory, Stackelberg game theory, and reachability checks.

[0114] In practical implementation, the first step is to model the maneuvering behavior of the UAV. When faced with abnormal navigation signals, UAVs typically execute preset strategies, such as hovering, acceleration for penetration, deceleration for hovering, climbing for escape, and return-to-home switching. These actions are abstracted into a finite set of actions, each with a clear dynamic manifestation. For example, hovering is represented by a circular trajectory of a fixed radius, acceleration for penetration is represented by a sudden increase in horizontal acceleration, and return-to-home switching is represented by a sharp turn in the flight path. By extracting and modeling these behaviors, game theory prediction can cover most of the countermeasures that the UAV may take.

[0115] After defining the action set, the system establishes a minimax game model. The decoy system aims to minimize the deviation between the generated trajectory and the UAV's actual trajectory, while the UAV aims to maximize the deviation or successfully return to base. This model allows for the numerical optimization of the equilibrium solution for both sides under the worst-case scenario. This equilibrium solution ensures that the decoy system maintains a certain level of trajectory prediction accuracy even in the worst-case scenario, enhancing its robustness.

[0116] To better align with the response logic of drones, this embodiment further introduces a Stackelberg leader-follower game. The deception system, acting as the leader, first sets a deception strategy; the drone, acting as the follower, selects its optimal maneuver under this strategy. By solving the Nash equilibrium of the leader-follower game, the set of drone response trajectories under different deception schemes can be obtained. This approach is more realistic than minimax game theory because drones mostly respond passively rather than actively.

[0117] After generating the set of adversarial trajectories, the system also needs to perform reachability checks on them. These checks include whether the trajectory is within the UAV's state space, whether the acceleration exceeds the maximum overload capacity, and whether the UAV has sufficient remaining energy to support the corresponding maneuver. This process eliminates physically impossible trajectories, improving the reliability of the prediction results.

[0118] Ultimately, the game theory predictions, after reachability testing, are fused with the dynamics and temporal predictions. If all three are consistent, the prediction is adopted directly; otherwise, the game theory predictions are used as constraints to correct the other predictions. This results in a prediction sequence that better reflects the actual reaction patterns of the UAV.

[0119] This implementation introduces adversarial logic into trajectory prediction, enabling the decoy system to counter UAV countermeasures. A finite action set ensures coverage of major maneuver types, minimax game theory enhances prediction robustness, Stackelberg game theory aligns with UAV response logic, and reachability checks ensure physical feasibility. Overall, this method addresses the lack of adversarial adaptability and physical verification in existing technologies, making predicted trajectories more natural and reliable in adversarial environments, thereby improving the success rate of soft-landing decoys.

[0120] In this embodiment, the consistency verification and constraint shaping include:

[0121] Error statistical models are established for GNSS, IMU, barometric and visual signals, respectively. The models include Gaussian noise, low-frequency bias and sudden interference.

[0122] Calculate the joint residual vector between different signal sources and compare it with a preset statistical threshold;

[0123] The residual results are input into the constraint optimizer, and a constraint solution is generated using quadratic programming or convex optimization methods.

[0124] The amplitude and timing of the virtual signal are adjusted according to the constraint solution to ensure that the multi-source signals maintain consistency in the statistical sense of covariance.

[0125] In this embodiment, the dynamic weight coordination and phased guidance include:

[0126] The flight process of the drone is divided into four stages: long-range high altitude, medium-altitude transition, low-altitude approach, and terminal landing.

[0127] During the long-range, high-altitude phase, the weight of GNSS signals shall not be less than 70%;

[0128] During the hollow transition phase, the weight of the air pressure signal shall be no less than 20%, and the weight of the IMU signal shall be no less than 25%.

[0129] During the low-altitude approach phase, the weight of visual signals shall be no less than 40%, and the weight of barometric signals shall be no less than 30%.

[0130] During the terminal landing phase, the policy network based on the Markov decision process adjusts the weights of various signals in real time.

[0131] The transition between each stage is managed by a finite state machine to achieve a smooth transition.

[0132] In this embodiment, the progressive perturbation control satisfies the following constraints:

[0133] The amplitude of a single-step perturbation, after statistical testing, does not exceed twice the standard deviation of the sensor noise.

[0134] The first-order difference and the second-order difference at adjacent time points are both less than the preset threshold;

[0135] The pseudorange, carrier phase, and code phase of GNSS signals are independently constrained in both the frequency and time domains;

[0136] The bias, scaling factor, and random walk components of the IMU signal are limited by amplitude and bandwidth, respectively.

[0137] Visual signal perturbations are generated through pixel affine transformation and illumination perturbation, and are constrained by the upper limit of projection error.

[0138] In this embodiment, the closed-loop adaptive adjustment includes:

[0139] Construct a feature vector based on the difference between the observed trajectory and the predicted trajectory;

[0140] The feature vector is input into a reinforcement learning agent, which employs a deep Q-network or a near-end policy optimization algorithm.

[0141] Based on the agent output, the parameters of the prediction module, signal generation module, and weight allocation module are updated in real time.

[0142] During the update process, a sliding window memory buffer and a priority experience replay mechanism are used.

[0143] When the drone exhibits abnormal maneuvers, the parameters are rolled back to the previous stable solution.

[0144] In this embodiment, the near-ground landing triggering and execution includes:

[0145] Establish a geometric model of the safe landing zone in the target area, the model including open areas, runways, or artificial landmark areas;

[0146] Landing trigger conditions are set based on the drone's altitude and speed thresholds, and the conditions include return-to-home, low battery, or mission completion logic.

[0147] After the triggering conditions are met, glide path parameters are generated, including heading alignment angle, descent rate curve and contact point coordinates;

[0148] During landing, adjust the drone's descent rate, throttle, and attitude input to conform to the glide path parameters.

[0149] Virtual visual markers or optical landmarks are set up in the contact point area to assist the drone in completing the grounding determination.

[0150] In this embodiment, the generation and transmission of the multi-source collaborative virtual signal includes:

[0151] The GNSS channel generates code phase, carrier phase, and Doppler frequency shift, and performs bandwidth shaping and power spectrum constraint.

[0152] A virtual atmospheric model of pressure channel variation with altitude is constructed, which includes temperature gradient, humidity correction and wind speed disturbance;

[0153] The IMU channel injects slowly varying bias, low-frequency random walk, and band-limited high-frequency noise into the triaxial gyroscope and accelerometer.

[0154] The visual channel generates a set of virtual optical landmarks and performs projection transformations based on camera parameters to obtain continuous virtual image frames;

[0155] The multi-channel signals are synchronized with a unified clock, including GPS clock alignment, hardware clock correction and delay compensation;

[0156] The alignment error of the multi-channel transmission is no greater than 1 microsecond and is compensated in real time through a phase-locked loop.

[0157] In one embodiment, a computer program product is also provided, including a computer program / instructions that, when executed by a processor, relate to all or part of the processes in the methods of the above embodiments.

[0158] The proposed scheme is limited to implementation by public administration departments (authorized by the government or granted power by law) or their designated units, and the specific implementation process shall comply with relevant laws and regulations.

[0159] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A soft landing control method for unmanned aerial vehicles (UAVs) based on trajectory prediction and multi-source cooperative deception, characterized in that, The method includes: Acquire multimodal state information of the target UAV and identify the model and flight control logic. The multimodal state information includes position, velocity, attitude, trajectory sequence, and navigation and anomaly detection features. The state information is used to perform hierarchical trajectory prediction, which includes short-term prediction based on a six-degree-of-freedom dynamic model, medium- and long-term prediction based on a temporal deep network, and game-theoretic prediction based on the adversarial maneuver hypothesis. The different prediction results are then fused to generate a decoy target corridor and descent profile. Based on the prediction results, a multi-source collaborative virtual signal is generated, which includes at least a satellite navigation virtual signal, a barometric altitude virtual quantity, an inertial measurement unit perturbation quantity, and a visual landmark reference. The multi-source collaborative virtual signal is subjected to consistency verification and constraint shaping to meet the requirements of spatiotemporal alignment, dynamic reachability and sensor noise statistical characteristics. The virtual signals are dynamically weighted according to the flight stage of the UAV, and a smooth transition is maintained when the weights change. A gradual perturbation is applied to the virtual signal to guide the drone to gradually deviate from its original trajectory and enter the decoy target corridor; Closed-loop adaptive adjustment of prediction parameters, signal generation parameters, and weight vectors is performed based on the real-time feedback status of the UAV. When the drone enters the preset corridor terminal, it provides its flight control system with landing trigger conditions and landing reference information, enabling the drone to complete a soft landing.

2. The method according to claim 1, characterized in that, The fusion method for the hierarchical trajectory prediction is as follows: By using dynamic prediction as the process model input for the extended Kalman filter, the prediction mean and covariance are established. The time series prediction results are used as the observation prediction input, and the residuals are compared with the dynamic prediction. When the residuals are lower than the preset threshold, an update is performed, and when they are higher than the threshold, a fault-tolerant update strategy is adopted. The game prediction results are used as a constraint set or penalty cost term to guide the final prediction trajectory; The covariance cross-update and Bayesian weight allocation methods are used to jointly optimize different prediction results.

3. The method according to claim 1, characterized in that, The game-theoretic prediction includes: The target UAV's maneuver strategy is modeled as a finite set of actions, which includes hovering, acceleration for penetration, deceleration for hovering, climbing for escape, and return-to-home switching. Establish a minimax game model so that the deception system takes minimizing trajectory deviation as its payoff function, while the drone takes increasing trajectory deviation or successfully returning to base as its payoff function. The Stackelberg leader-follower game model is adopted, with the decoy system as the leader and the drone as the follower. The set of adversarial trajectories is generated by solving the game equilibrium. The reachability of the set of adversarial trajectories is checked, including constraints on state space boundaries, acceleration limits, and energy margins. The trajectory envelope that passes the test is cross-compared with other prediction results to obtain the corrected game prediction sequence.

4. The method according to claim 1, characterized in that, The consistency verification and constraint shaping include: Error statistical models were established for GNSS, IMU, barometric and visual signals, respectively. The models included Gaussian noise, low-frequency bias and sudden interference. Calculate the joint residual vector between different signal sources and compare it with a preset statistical threshold; The residual results are input into the constraint optimizer, and a constraint solution is generated using quadratic programming or convex optimization methods. The amplitude and timing of the virtual signal are adjusted based on the constraint solution to ensure that the multi-source signals maintain consistency in the statistical sense of covariance.

5. The method according to claim 1, characterized in that, The dynamic weight coordination and phased guidance include: The flight process of the drone is divided into four stages: long-range high altitude, medium-altitude transition, low-altitude approach, and terminal landing. During the long-range, high-altitude phase, the weight of GNSS signals shall not be less than 70%; During the hollow transition phase, the weight of the air pressure signal shall be no less than 20%, and the weight of the IMU signal shall be no less than 25%. During the low-altitude approach phase, the weight of visual signals shall be no less than 40%, and the weight of barometric signals shall be no less than 30%. During the terminal landing phase, the policy network based on the Markov decision process adjusts the weights of various signals in real time. The transition between each stage is managed by a finite state machine to achieve a smooth transition.

6. The method according to claim 1, characterized in that, The progressive perturbation control satisfies the following constraints: The amplitude of a single-step perturbation, after statistical testing, does not exceed twice the standard deviation of the sensor noise. The first-order difference and the second-order difference at adjacent time points are both less than the preset threshold; The pseudorange, carrier phase, and code phase of GNSS signals are independently constrained in both the frequency and time domains; The bias, scaling factor, and random walk components of the IMU signal are limited by amplitude and bandwidth, respectively. Visual signal perturbations are generated through pixel affine transformation and illumination perturbation, and are constrained by the upper limit of projection error.

7. The method according to claim 1, characterized in that, The closed-loop adaptive adjustment includes: Construct a feature vector based on the difference between the observed trajectory and the predicted trajectory; The feature vector is input into a reinforcement learning agent, which employs a deep Q-network or a near-end policy optimization algorithm. Based on the agent output, the parameters of the prediction module, signal generation module, and weight allocation module are updated in real time. During the update process, a sliding window memory buffer and a priority experience replay mechanism are used. When the drone exhibits abnormal maneuvers, the parameters are rolled back to the previous stable solution.

8. The method according to claim 1, characterized in that, The method further includes: Near-ground landing triggering and execution: Establish a geometric model of the safe landing zone in the target area, including open areas, runways, or artificial landmarks; Set landing trigger conditions based on drone altitude and speed thresholds, including return-to-home, low battery, or mission completion logic. After the triggering conditions are met, glide path parameters are generated, including heading alignment angle, descent rate curve and contact point coordinates. During landing, adjust the drone's descent rate, throttle, and attitude input to conform to the glide path parameters. Virtual visual markers or optical landmarks are set up in the contact point area to assist the drone in completing the grounding determination.

9. The method according to claim 1, characterized in that, The generation and transmission of the multi-source collaborative virtual signal includes: The GNSS channel generates code phase, carrier phase, and Doppler frequency shift, and performs bandwidth shaping and power spectrum constraint. A virtual atmospheric model of pressure channel variation with altitude is constructed, which includes temperature gradient, humidity correction and wind speed disturbance; The IMU channel injects slowly varying bias, low-frequency random walk, and band-limited high-frequency noise into the triaxial gyroscope and accelerometer. The visual channel generates a set of virtual optical landmarks and performs projection transformations based on camera parameters to obtain continuous virtual image frames; Multi-channel signals are synchronized with a unified clock, including GPS clock alignment, hardware clock correction and delay compensation; The alignment error of multi-channel transmission is no greater than 1 microsecond, and is compensated in real time through a phase-locked loop.

10. A computer program product, characterized in that, Includes a computer program / instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1-9.

Images