A Resilient Control Method and System for Brain-Controlled Unmanned Aerial Vehicles Based on a Large Model

Abstract

This invention discloses a robust control method and system for brain-controlled unmanned aerial vehicles (UAVs) based on a large model, belonging to the field of brain-computer interface and UAV control technology. Specifically, it includes: acquiring EEG signals and identifying discrete brain-controlled event commands; inputting these commands into a large model intent reconstruction module after sliding window serialization encoding; outputting action candidates with multiple risk levels; performing parsing verification, risk assessment, and motion feasibility checks on the action candidates; selecting the final action and mapping it to parameterized motion skill atoms; having a state machine manage the execution of the skill atoms; implementing closed-loop execution via the flight control interface and monitoring safety boundaries; and triggering hovering and return-to-home actions in case of anomalies. This invention achieves intent reconstruction and risk-layered decision-making for brain-controlled commands, constructing a full-link safety closed-loop control system. It significantly improves the control stability, safety, and system resilience of brain-controlled UAVs under uncertain command conditions, reduces the risk of false triggering, and increases the mission success rate.

Description

Technical Field

[0001] This invention belongs to the field of brain-computer interface and unmanned aerial vehicle (UAV) control technology, specifically relating to a resilient control method and system for brain-controlled UAVs based on a large model. Background Technology

[0002] Patent 201710089650.1 discloses an Emotiv brain-controlled drone system and method based on hybrid programming of VC++ and Matlab. The system includes: an Emotiv EEG signal acquisition module, a Matlab EEG signal online analysis module, and a drone flight control module. The Emotiv EEG signal acquisition module acquires the brain signals of the subject controlling the drone through an Emotiv device. The brain signals are transmitted to VC++ acquisition software on a computer, and the acquisition software transmits them to the Matlab EEG signal online analysis module via TCP / IP protocol. The acquired EEG signals are preprocessed, feature extracted, classified, and analyzed to identify effective components. These are then converted into control commands and sent to the drone flight control module via a wireless serial port. Upon receiving the control commands, the drone executes takeoff, landing, left-flying, and right-flying flight operations. The use of hybrid programming of VC++ and Matlab allows for easy embedding of brain-computer interface pattern recognition methods and practical hardware and software applications, laying the foundation for online analysis and practical applications of brain-computer interfaces. However, the problem with this method is that the classified output is directly mapped to actions, lacking system-level processing for consistency of historical instruction sequences, delay conflicts, and false triggers; airborne execution often focuses on executing whatever is received, lacking closed-loop safety safeguards and verification of action rationality.

[0003] Patent CN108415565A discloses a brain-computer interface (BCI) intelligent control method and technology for unmanned aerial vehicles (UAVs). This BCI system includes a visual stimulation induction module, a real-time image transmission module, an EEG acquisition device, a multi-feature EEG processing module, a brain-computer communication module, and a UAV and onboard microcontroller module. The experimenter can control various actions of the UAV without any physical movements or language. The EEG of the experimenter is collected using an EEG detector; the EEG features include steady-state visual stimulation and imagined movement, switched by blinking. The EEG is preprocessed, feature extracted, and classified in the processing module, then converted into control commands. These commands are sent to the UAV's onboard microcontroller module via the communication system. Simultaneously, the experimenter observes the UAV's flight through a global camera, plans the flight route, and ultimately controls the UAV to complete the preset flight mission. This invention constructs a multi-mode BCI system, enabling flexible direct brain control of the UAV and effectively compensating for the shortcomings of existing UAV control methods. The shortcomings of this scheme are: in order to perform more actions, the instruction set is expanded by multi-paradigm fusion or switching; the essence is still the "identification → discrete command → direct control" paradigm, which emphasizes "controllability", but the handling of "stability and safety caused by unreliable commands (noise, delay, false triggering / omission coexistence)" is still mainly based on simple rules; safety constraints usually exist as inherent capabilities of the flight control side or human operation experience, and lack a strategic mechanism of "risk stratification + backoff" for joint modeling of brain control uncertainty.

[0004] Patent CN105549743A discloses a robot system and its implementation method based on a brain-computer interface (BCI). Its main features include: a BCI subsystem acquiring electroencephalogram (EEG) signals through measuring electrodes and extracting features from the EEG signals to obtain the user's intention; the BCI subsystem converting the user's intention into control commands and transmitting them to the robot subsystem via a communication subsystem; the robot subsystem receiving the control commands from the BCI subsystem and controlling the robot's movement; and the robot subsystem transmitting its own collected video information to the BCI subsystem, which then generates a stimulation interface to provide feedback to the user. This achieves steady-state visual evoked potential (VEVP) BCI control of the robot's movement in eight directions, and the system can achieve rapid and precise control of the robot without training. Since the BCI does not require limb movement, it can move users with severe motor impairments to the desired location, improving their quality of life. The advantage of this approach lies in providing a closed-loop interaction pattern of "perceptual feedback—stimulation interface—intention control" and enabling multi-directional movement control. However, the main drawback is that the control chain still primarily relies on discrete intentions to directly drive movement, and it does not provide a safety-oriented, resilient control framework to address the conflicts between false triggering and delays in brain-controlled input. As can be seen from the above solutions, the main shortcoming of existing brain-controlled drone / robot control systems lies not in their ability to decode intentions, but in how to ensure flight stability and safety when the decoding results are unreliable. The main drawbacks directly related to the problem this invention aims to solve include: (1) Direct mapping control of discrete commands lacks resilience in handling the uncertainty of brain control. Existing solutions generally adopt the control method of "identifying and outputting a discrete command → executing it immediately". When there is noise, drift, short-term distortion or fluctuation in recognition confidence in the EEG, false triggering, omission, repeated triggering and delayed triggering are likely to occur, resulting in the UAV's jittery movements, frequent changes of direction or performing dangerous actions at inappropriate times (such as accelerating forward when approaching obstacles).

[0005] (2) Lack of an intent reconstruction mechanism based on "historical sequence consistency / contextual reasoning". Existing solutions are mostly based on single classification output or short-term majority voting. Such methods can only perform simple smoothing and are difficult to use "event sequence within the sliding window + flight state context" to infer more reasonable and executable true intents (e.g., continuous target intents when recognizing the "left / left / front" combination).

[0006] (3) Lack of risk-stratified candidate action generation and rollback strategies. Existing solutions typically output a single action instruction and lack a mechanism to simultaneously generate multi-level strategy candidates of "safe / compromise / aggressive" and automatically select and roll back based on confidence level, environmental risk and constraints. Therefore, when the environment is complex or the input is uncertain, it is impossible to balance safety and efficiency.

[0007] (4) Lack of a unified closed loop for hard constraint feasibility verification and safety fallback. Although the flight control itself may have the ability to return to home / land, the existing solutions often fail to incorporate factors such as "insufficient confidence, insufficient obstacle clearance, fence triggering, low battery, link anomaly" and the uncertainty of brain control decoding into a unified decision-making closed loop, resulting in delayed triggering of safety strategies or scattered logic. Summary of the Invention

[0008] The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing a resilient control method and system for brain-controlled drones based on a large model. This method enables drones to output stable, safe, and executable control behaviors even when faced with discrete, unreliable, delayed, and noisy brain-controlled commands, thereby improving system resilience and reliability. It addresses the technical problems in existing brain-controlled drone control systems, such as poor anti-interference capability due to direct mapping of discrete commands, lack of intent reconstruction mechanism based on historical sequence context reasoning, lack of risk-free hierarchical candidate action generation and backoff strategies, and lack of a unified closed loop for hard constraint feasibility verification and safety fallback. These problems lead to drone jitter, high risk of false triggering, poor flight stability and safety, and an inability to stably perform control tasks in complex brain-controlled input and flight environments.

[0009] The present invention adopts the following technical solution: A robust control method for brain-controlled unmanned aerial vehicles based on a large model includes the following steps: S1. Collect and identify the user's brainwave signals to obtain discrete brain control event instructions, wherein the discrete brain control event instructions are accompanied by at least identification confidence and / or signal quality indicators. S2. Within a sliding window of a preset time length, the current discrete mind control event command and the historical discrete mind control event command are serialized and encoded to form a command sequence, and the command sequence is encoded into a symbol sequence that can be processed by a large model. S3. Input the symbol sequence and the UAV state into the large model intent reconstruction module, and the large model intent reconstruction module outputs at least one action candidate and the corresponding candidate confidence. S4. Perform risk assessment and hierarchical decision-making on the candidate actions, and select the final action based on the confidence level of the candidates and the risk assessment results; S5. Perform a motion feasibility check on the final action; if the feasibility check fails or the candidate confidence level is lower than the threshold, revert to the action candidate under a lower risk strategy or trigger a protection strategy. S6. The final action that passes the feasibility check is mapped to one or more predefined parameterized motion skill atoms, and the state machine manages the entry conditions, execution process and exit conditions of the motion skill atoms. S7. The motion skill atoms are executed in a closed loop through the UAV flight control and flight control interface, and safety boundary monitoring is performed during the execution process; when the preset protection trigger conditions are met, one or more protection strategies are executed, such as hovering, slow descent, return to home or landing.

[0010] Preferably, in step S1, the EEG signal is preprocessed after acquisition, and the preprocessing includes at least one or more of bandpass filtering, power frequency suppression, artifact suppression, and baseline correction; the discretized brain-controlled event command is obtained based on the SSVEP paradigm and / or P300 paradigm, and the event type of the discretized brain-controlled event command includes at least one or more of takeoff, landing, hovering, forward, backward, leftward, rightward, leftward, rightward, ascent, descent, and return.

[0011] Preferably, in step S2, the length of the sliding window can be adaptively adjusted according to the signal quality index. When the signal quality decreases or the drift increases, the window length is increased; when the signal quality improves and the response requirement is higher, the window length is decreased. The symbol sequence is organized in the form of structured fields, including at least one or more of the following: event sequence field, UAV status field, and safety constraint field. Each discrete brain control event instruction in the instruction sequence includes at least two or more of the following: event type, timestamp, recognition confidence, and signal quality index.

[0012] Preferably, in step S3, the drone status includes at least one or more of the following: flight mode, altitude, speed vector, battery level, fence status, minimum obstacle distance, and communication link status; the action candidates are output in the form of structured fields, including at least one or more of the following: risk level field, action type field, parameter field, and candidate confidence field; the action candidates include at least a candidate set corresponding to two or three levels of risk strategies among safe candidates, compromise candidates, and aggressive candidates.

[0013] Preferably, in step S4, before performing risk assessment and hierarchical decision-making, the structured fields output by the large model intent reconstruction module are parsed and verified. The parsing and verification include at least field integrity checks, parameter range pruning, and / or default filling. When parsing or verification fails, the action candidate is not issued and the protection strategy is triggered or the action candidate is rolled back to a lower risk strategy. The risk assessment is based on at least two or more of the following constraints: maximum speed constraint, minimum obstacle distance constraint, geofence constraint, and low battery threshold constraint.

[0014] Preferably, in step S5, the motion feasibility check is based at least on one or more of the following: dynamic boundary constraints, flight mode executability constraints, fence boundary constraints, and obstacle distance constraints, to determine whether the final action or skill atomic parameters are executable; the protection strategy includes at least one or more of the following: entering hovering mode, return-to-home mode, and / or landing mode.

[0015] Preferably, in step S6, the motion skill atom includes at least one or more of the following: takeoff, landing, hovering, straight movement, backward movement, left movement, right movement, yaw rotation, ascent, descent, and return. Each skill atom contains one or more adjustable parameters of speed, angular velocity, duration, target altitude, or turning angle. The state machine includes at least an idle state, a skill execution state, a safety protection state, and a landing state, and supports transitioning from any state to the safety protection state under abnormal conditions.

[0016] Preferably, during the skill execution state, new discrete mind control event commands are locked, suppressed, or delayed; the locking process is to ignore the newly entered discrete mind control event command, the suppression process is to only allow high-priority safety event commands to interrupt the execution of the current skill atom, and the delay process is to cache the new discrete mind control event command until the current skill atom is completed before participating in the next round of intent reconstruction; the high-priority safety event includes at least one or more of hovering, returning to home, and / or landing.

[0017] Preferably, in step S7, the protection triggering conditions include at least: candidate confidence level below a threshold, obstacle distance less than the minimum obstacle distance threshold, triggering fence boundary, low battery threshold triggering, communication anomaly, flight status anomaly, and large model output parsing failure or verification failure; the control commands issued by the flight control interface include at least one or more of speed / heading control commands, mode switching commands, and task / action commands; when the large model is unavailable or the communication link is poor, the system enters a degraded mode, using a rule engine / finite state table to replace the large model intent reconstruction, allowing only safety skills to be executed atomically and increasing the confidence threshold and risk level.

[0018] Secondly, embodiments of the present invention provide a robust control system for brain-controlled unmanned aerial vehicles based on a large model, comprising: The recognition module is used to collect and recognize the user's brain signals and output discrete brain control event commands, wherein the discrete brain control event commands are accompanied by at least recognition confidence and / or signal quality indicators. The encoding module is used to serialize and encode the current discrete mind control event command and the historical discrete mind control event command within a sliding window of a preset time length, forming a command sequence, and then encode the command sequence into a symbol sequence that can be processed by a large model. The reconstruction module is used to receive the symbol sequence and the UAV state, and output at least one action candidate and the corresponding candidate confidence level; The decision module is used to perform risk assessment and hierarchical decision-making on the action candidates, select the final action based on the confidence level of the candidates and the risk assessment results, and parse and verify the structured fields output by the large model before making the decision. The inspection module is used to perform a motion feasibility check on the final action. When the feasibility check fails or the candidate confidence level is lower than the threshold, a rollback mechanism or protection strategy is triggered. The management module is used to map the final action that passes the feasibility check to one or more predefined parameterized motion skill atoms, and the state machine manages the entry conditions, execution process and exit conditions of the motion skill atoms. The monitoring module is used to execute the motion skill atoms through the UAV flight control and closed loop via the flight control interface. During the execution, it performs safety boundary monitoring and executes one or more protection strategies such as hovering, slow descent, return to home, or landing when the preset protection trigger conditions are met.

[0019] Thirdly, a computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described brain-controlled unmanned aerial vehicle resilience control method based on a large model.

[0020] Fourthly, embodiments of the present invention provide a computer-readable storage medium including a computer program, which, when executed by a processor, implements the steps of the above-described brain-controlled unmanned aerial vehicle resilience control method based on a large model.

[0021] Fifthly, a chip includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described brain-controlled unmanned aerial vehicle resilience control method based on a large model.

[0022] Sixthly, embodiments of the present invention provide an electronic device including a computer program, which, when executed by the electronic device, implements the steps of the above-described brain-controlled drone resilience control method based on a large model.

[0023] Compared with the prior art, the present invention has at least the following beneficial effects: A robust control method for brain-controlled unmanned aerial vehicles (UAVs) based on a large model is proposed. Taking discrete brain-controlled signal events as input and executable motor skill atomic sequences as output, it forms a closed-loop control chain encompassing brain-controlled signal recognition, serialized token encoding, intent reconstruction, risk-level decision-making, motor feasibility checking, motor skill atomic execution, and safety fallback. This constructs a complete technical chain for robust control of brain-controlled UAVs based on a large model, forming a closed-loop process from EEG signal acquisition to final flight control execution and safety protection. It combines large-model intent reconstruction, risk-level decision-making, and motor skill atomic execution, overcoming the shortcomings of existing technologies that directly map discrete commands. Uncertainty handling of brain-controlled commands is integrated throughout the entire process. Through serialization encoding, intent reconstruction, feasibility checking, and state machine management, noise, delay, and false triggering issues in brain-controlled signals are filtered layer by layer. Simultaneously, a unified protection strategy triggering mechanism is established, ensuring the UAV remains stable and controllable even under unreliable brain-controlled input. Compared to the single command execution logic of existing technologies, this method forms a robust control system encompassing recognition, reconstruction, decision-making, execution, and protection.

[0024] Furthermore, preprocessing techniques such as bandpass filtering effectively remove signal noise, power frequency interference, and artifacts, improving the effectiveness of EEG signals. The recognition method based on the SSVEP / P300 paradigm ensures the professionalism and accuracy of discrete command recognition, while the accompanying confidence and signal quality indicators provide data support for adaptive adjustments in subsequent stages. A rich variety of event types cover all UAV flight maneuvers, meeting practical control requirements. The EEG signal acquisition and recognition process has been refined and optimized, clarifying the EEG signal preprocessing methods and the recognition paradigm and type range of discrete brain-controlled event commands, solving the problems of low EEG signal recognition accuracy, limited command types, and lack of quality characterization in existing technologies.

[0025] Furthermore, the sliding window length adaptively adjusts with signal quality, increasing the window to improve stability when the signal is poor and decreasing the window to ensure response speed when the signal is good, thus balancing control stability and real-time performance. The instruction sequence contains multi-dimensional event information, allowing the large model to acquire more comprehensive historical and current instruction data. The structured symbol sequence organization conforms to the processing logic of the large model, reducing encoding ambiguity and improving the accuracy of intent reconstruction. The serialization encoding process has been optimized, achieving adaptive adjustment of the sliding window and structured organization of the symbol sequence, solving the intent inference bias problems caused by the lack of historical sequence analysis and non-standard encoding formats in existing technologies.

[0026] Furthermore, multi-dimensional acquisition of UAV status allows the large model to perform contextual reasoning in conjunction with the flight environment, avoiding unrealistic intent reconstruction; structured action candidate output allows for direct parsing and processing in subsequent decision-making stages, improving engineering feasibility; and the risk-layered candidate structure of safety / compromise / aggressive actions allows the system to flexibly select actions based on actual conditions, balancing flight safety and operational efficiency. This clarifies the input and output specifications for the intent reconstruction stage of the large model, enriches the dimensions of UAV status monitoring and the risk-layered form of action candidates, and solves the problems of existing technologies having no context for intent inference, single output actions, and no risk differentiation.

[0027] Furthermore, the parsing and verification step performs multi-dimensional checks on the output of the large model, filters out abnormal outputs, and prevents invalid or dangerous actions from entering the execution stage, thus adding a security barrier to the control link. Risk assessment based on multi-dimensional security constraints provides clear quantitative basis for the decision-making process, avoiding subjective judgment. The rollback and protection mechanism after parsing failure achieves early interception of risks. The parsing and verification step is added to the risk assessment and hierarchical decision-making stages, and the core constraints of risk assessment are clarified, solving the problems of existing technologies lacking large model output verification, lacking clear basis for risk assessment, and having simple decision-making logic.

[0028] Furthermore, multi-dimensional feasibility checks and constraints verify the feasibility of actions from aspects such as dynamics, flight modes, and environmental boundaries, avoiding actions that exceed the performance or environmental limitations of the UAV and eliminating dangerous flight behaviors at the source. The clear protection strategy allows the system to quickly make standardized safety responses when the feasibility check fails, avoiding helpless or erroneous emergency operations. The judgment criteria for the motion feasibility check are refined, and the specific form of the protection strategy is clarified, solving the problem of high execution risks caused by the lack of motion feasibility verification and the ambiguity of protection strategies in existing technologies.

[0029] Furthermore, parameterized motion skill atoms decompose the complex movements of drones into standardized, autonomously executable basic units, decoupling brain-controlled triggering from autonomous drone execution and reducing reliance on continuous, high-quality brain-controlled signals. Multi-state state machine management ensures clear entry, execution, and exit rules for the execution process of skill atoms, while a rapid transfer mechanism in abnormal states guarantees the stability of the execution process, solving the problems of direct instruction execution, lack of standardized motion units, and lack of state management in existing technologies.

[0030] Furthermore, the input processing mechanism of locking / suppressing / delaying effectively avoids interference from irrelevant brain-controlled commands during the execution of skills atoms, ensuring the continuity of action execution; it only allows high-priority safety events to interrupt execution, balancing execution continuity with emergency safety, and avoiding action chaos caused by unnecessary command interruptions; the clear rules for different processing methods make input management more standardized, improve the engineering feasibility of the system, clarify the rules for different processing methods and the scope of high-priority safety events, and solve the problem of flight jitter caused by unconstrained intervention of new commands and easy command conflicts during the execution of existing technical commands.

[0031] Furthermore, the multi-dimensional protection trigger conditions enable comprehensive monitoring of risks across the entire link, allowing the system to trigger protection in a timely manner when an anomaly occurs at any stage; the standardized flight control command types conform to the engineering interface logic of UAV flight control, improving the accuracy of command execution; the degradation mode allows the system to maintain basic safety control capabilities even when there are large model or communication anomalies, avoiding complete loss of system control. The flight control command types are clearly defined, and a system degradation mode is added, solving the problems of existing technologies having single protection trigger conditions, non-standard flight control commands, and no extreme situation response mechanism.

[0032] It is understood that the beneficial effects of the second to sixth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0033] In summary, the method of the present invention can significantly improve the stability, safety and robustness of UAV control under conditions such as discrete brain control commands, high noise, time delay fluctuations, and the coexistence of false triggering and omission.

[0034] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a diagram showing the state machine for the execution of the skill primitives and the state machine for transitioning between abnormal states in this invention. Figure 3 This is a diagram illustrating risk assessment and tiered decision-making. Figure 4 A schematic diagram of a computer device provided in an embodiment of the present invention; Figure 5 This is a block diagram of a chip provided according to an embodiment of the present invention.

[0036] Among them, 60. Computer equipment; 61. Processor; 62. Memory; 63. Computer program; 600. Electronic device; 610. Processing unit; 620. Storage unit; 6201. Random access memory unit; 6202. Cache memory unit; 6203. Read-only memory unit; 6204. Program / utility; 6205. Program module; 630. Bus; 640. Display unit; 650. Input / output interface; 660. Network adapter; 700. External device. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0039] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0040] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Additionally, the character " / " in this invention generally indicates that the preceding and following objects have an "or" relationship.

[0041] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0042] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0043] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0044] This invention provides a resilient control method for brain-controlled unmanned aerial vehicles (UAVs) based on a large model, relating to the fields of brain-computer interface (BCI) and UAV control technology. Specifically, it relates to a resilient control method for brain-controlled UAVs that utilizes a large model to reconstruct intent and make risk decisions for discrete, noisy, delayed, and unreliable brain-controlled commands, combined with local closed-loop autonomous execution and safety fallback of motor skill atoms. The method includes: acquiring EEG signals and identifying discrete brain-controlled event commands; serializing and encoding current and historical events into a command sequence within a sliding time window; inputting the command sequence into a large model intent reconstruction module to output action candidates and their confidence levels; performing risk assessment and hierarchical decision-making on the action candidates, and selecting the final action based on motion feasibility checks; mapping the final action to one or more parameterized motor skill atoms, which are then controlled by a state machine. The system manages the entry, execution, and exit of skill atoms; it executes skill atoms in a closed loop via flight control and its interface, and monitors safety boundaries during execution. When confidence is insufficient, feasibility is insufficient, or flight status is abnormal, it triggers protection strategies such as hovering, slow descent, return to home, or landing. It introduces large-model intent reconstruction and risk-layered rollback, and forms an engineering closed loop through flight control-executable LOITER / RTL / LAND protection strategies. It uses a large model to reason about event sequences and flight status, handling conflicting commands and contextual changes, and combining fences, obstacle departure, and power thresholds for safety decisions. Compared to shared control without large-model reasoning, this invention uses intent reconstruction, risk layering, and feasibility verification as a resilience hub, making skill atom triggering and execution more stable and reducing the risk of false triggering.

[0045] The key data is defined as follows: (1) Discretized brain control signal event commands: event objects output by the brain control signal recognition module i Recorded as The event object contains at least two or more of the following fields: Event type, including at least one or more of the following: {TAKEOFF, LAND, HOVER, FORWARD, BACKWARD, LEFT, RIGHT, YAW_L, YAW_R, UP, DOWN, RTL}; : timestamp; : Identify confidence level, with a value range of ([0, 1]); Signal quality metrics are used to characterize channel availability, artifact rate, stability, or drift level, etc.

[0046] (2) Drone Status: The drone status is denoted as (S) and includes at least one or more of the following fields: flight mode, such as GUIDED, LOITER, RTL, LAND, etc.; altitude h; velocity vector v; battery; fence status (whether the horizontal / height fence is triggered or approaching); minimum obstacle distance. (Obtained by ranging methods such as vision / LiDAR / ultrasound), and communication link status (whether the heartbeat is normal, number of timeouts, etc.).

[0047] (3) Action Candidates and Motor Skill Atoms: The action candidates output by the large model are denoted as ( It must at least include the action type and parameter set (e.g., velocity v, angular velocity, etc.). ), duration t, target height h, turning angle ( (etc.); the action candidates are mapped to form motion skill atoms ( The motion skill atom is the basic flight behavior unit that the UAV can execute in a closed loop. It can rely on the UAV's onboard sensors to execute a locally autonomous trajectory segment, with clear entry / execution / exit conditions and parameterized control quantities.

[0048] Please see Figure 1 This invention discloses a resilient control method for brain-controlled unmanned aerial vehicles based on a large model, comprising the following steps: S1. Collect and identify the user's EEG signal to obtain discrete brain control event commands, and add identification confidence and / or signal quality indicators to the discrete brain control event commands; This invention acquires electroencephalogram (EEG) signals and identifies discrete brain-controlled event commands based on paradigms such as SSVEP and / or P300. The acquired EEG signals undergo preprocessing, which includes at least one or more of bandpass filtering, power line suppression, artifact suppression, and baseline correction.

[0049] In one implementation, the SSVEP paradigm is employed: frequency domain features are extracted from multi-frequency stimulus targets, and correlation analysis or a classifier is used to output the target category, thereby obtaining the event type. and identification confidence level .

[0050] In another implementation, the P300 paradigm is used: ERP detection and classification are performed under the novel stimulus (Oddball) or similar event-related potential (ERP) evoked paradigm, the target selected event is output, and the event type is obtained. and identification confidence level .

[0051] Furthermore, the identification module can output signal quality indicators. Examples of metrics include channel disconnection rate, artifact ratio, and short-term stability indicators, which can be used to adaptively adjust the sliding window length, confidence threshold, and risk strategy selection.

[0052] S2. Within a sliding window of a preset time length, the current event and historical events are serialized and encoded to form an instruction sequence, and the instruction sequence is encoded into a symbol sequence (token) that can be processed by a large model; preferably, the symbol sequence (token) is organized in the form of structured fields, including at least one or more of the following: event sequence field, UAV status field, and / or safety constraint field; Each event within the sliding window must include at least two or more of the following: event type, timestamp, identification confidence level, and signal quality metrics.

[0053] Preferably, the event sequence is represented as a structured field event, wherein each event includes one or more of event type, timestamp, confidence level, and quality.

[0054] Regarding the sliding window construction and encoding method, a sliding time window (W) of length (T) is set, and the discrete event objects within the window are sorted by time to form an event sequence:

[0055] The window length (T) can be set to a range (e.g., 1s to 5s) and can be adaptively adjusted according to the signal indicators: when the quality deteriorates or the drift increases, the window length is increased to enhance stability; when the quality improves and the response requirements are higher, the window length is decreased to improve the response speed.

[0056] Event sequence ( The input (X) is encoded into a format that can be processed by a large model. Encoding methods include, but are not limited to: (1) Token sequence encoding: Mapping event types to discrete tokens; mapping event intervals to discrete tokens. Interval discretization is performed to obtain time markers; confidence level (conf) is discretized into intervals to obtain confidence markers; quality is mapped to grade markers; and these markers are then concatenated in chronological order to form a marker sequence.

[0057] (2) Structured encoding: Input the event sequence as a key-value pair sequence or JSON structure to reduce ambiguity and facilitate parsing and verification.

[0058] Preferably, structured coding is used to ensure the controllability and verifiability of the input and output of large models.

[0059] Existing technologies mostly generate instructions and execute them directly through single classification or simple voting; this invention serializes and encodes discrete events within a sliding window, inputs them into a large model for intent reconstruction, and uses historical event sequences and UAV state context to infer a more reasonable true intent, thereby reducing flight maneuver jitter and discontinuity caused by false triggers, omissions and delay conflicts, improving long-term control stability, and significantly enhancing the fault tolerance to brain-controlled noise / delay / drift.

[0060] Existing technologies lack structured constraints for large model outputs. This invention introduces a structured input / output field list for large models and performs field integrity checks, parameter range pruning, and default filling on the output. When parsing or verification fails, it directly enters protection / rollback mode, thereby preventing abnormal outputs from directly entering the flight control system and causing potential dangers. This mechanism allows large models to be engineered and safely introduced into the control link.

[0061] S3. Input the symbol sequence (token) into the large model intent reconstruction module, and the large model intent reconstruction module outputs at least one action candidate and its corresponding confidence level; Preferably, the action candidates are output in the form of structured fields, including at least a candidate list field and a candidate set, wherein each candidate includes one or more of a risk level field, an action type field, a parameter field, and a candidate confidence field.

[0062] This invention inputs the encoded input (X) and the UAV state (S) into a large model intent reconstruction module, and outputs a set of action candidates (X, X, S, ...). ) and its confidence set ( ), where each candidate action ( ) Attached candidate confidence ( ).

[0063] In one implementation, the large model outputs a candidate set under at least two or three levels of risk policies, such as safety candidates. ), compromise candidate ( ) and radical candidates ( ).

[0064] To ensure the feasibility and controllability of the project, this invention preferably uses structured input templates and structured output fields, and the parser performs field validation and range clipping.

[0065] (1) Input a list of structured fields for the large model (below is an example of an implementation) Input may be in JSON / key-value pair structure and must contain at least one or more of the following fields (fields may be truncated and are not required): events: A list of event sequences, each event containing: type: Event type (e.g., FORWARD / YAW_L / HOVER, etc.) dt: The time interval between the previous event (either a numerical value or a bucket label) conf: Identify confidence level quality: Signal quality level or numerical value uav_state: The drone state object, containing: mode: Flight mode (GUIDED / LOITER / RTL / LAND, etc.) h: height v: Current velocity (can be a scalar or vector magnitude) battery: percentage of charge or voltage level fence: Fence status (normal / approaching / triggered) d_obs: Minimum obstacle distance Link: Link health (normal / timeout count / packet loss rate, etc.) constraints: A safety constraint object containing: v_max: Maximum speed limit d_min: Minimum obstacle departure threshold B_low: Low battery threshold fence_rule: Fence rules (horizontal / height boundaries) goal / context (optional): Mission context (e.g., stay hovering, move forward through the doorway, return to the takeoff point, etc.).

[0066] (2) List of structured fields output by the large model Large model outputs can be structured lists, containing at least the following fields: candidates: A list of candidate actions, each candidate containing: policy_level: Risk level (safe / mid / aggr) action_type: Action type (e.g., HOVER / FORWARD / YAW / RTL / LAND, etc.) params: A set of parameters (e.g., v, ...). ,t,h, wait) confidence: candidate confidence level ) recommendation: Index or identifier of the recommended candidate (optional).

[0067] (3) Analysis and Verification The output is validated by the parser: field integrity check, numeric type check, and parameter range pruning (e.g., speed not exceeding a certain limit). ), default parameter filling. If parsing or verification fails, the candidate action will not be issued, and the protection strategy will be triggered directly or the candidate set of a more conservative strategy will be rolled back.

[0068] S4. Perform risk assessment and hierarchical decision-making on the candidate actions, and select the final action based on the confidence level and risk assessment results; After step S3, before step S4, and during step S4, the parsing and verification of the structured output of the large model are also included. The parsing and verification include at least field integrity checks, parameter range pruning, and / or default filling. When parsing or verification fails, the action candidate is not issued and the protection strategy is not triggered, or the action candidate under a lower risk strategy is rolled back. The protection strategy includes at least one or more of the following: LOITER (hover hold mode), RTL (return to home mode), and / or LAND (landing mode).

[0069] Please see Figure 3 It demonstrates a risk-based decision-making mechanism and a fallback mechanism, which are used to select safe and executable operations from multiple candidate solutions generated by a large language model. If a candidate action fails the risk or action feasibility check, the system will fall back to the next level of candidate solutions.

[0070] The large model output includes a set of action candidates corresponding to the safety strategy, compromise strategy and / or aggressive strategy, from which the final action is selected based on the environmental risk level and confidence level.

[0071] Risk assessment and tiered decision-making are based on at least two or more of the following safety constraints: maximum speed constraint, minimum obstacle distance constraint, geofence constraint, and low battery threshold constraint. Preferably, the safety constraints are input as structured fields into the large model intent reconstruction module and / or used for risk assessment and motion feasibility checks in steps S4 and S5.

[0072] Risk assessments are conducted on candidate actions, and decisions are made in a tiered manner to arrive at the final action. Risk assessments must be based on at least two or more of the following security constraints: (1) Maximum speed constraint: ); (2) Minimum obstacle distance constraint: ); (3) Fence constraints: Geographic fence or virtual fence (horizontal and height boundaries), restricting actions when the fence is triggered or approached; (4) Low battery threshold: When the time is limited, the aircraft will be allowed to return to base or land.

[0073] In one implementation, the risk assessment outputs a risk level (L) or a risk score ( And adopt the following exemplary rules (configurable, not limiting): (1) If ( If the condition is met, then RTL or LAND should be preferred, and aggressive candidates should not be selected. (2) If ( If a fence is triggered / approached, fall back to a safe candidate or perform a hover; (3) If the candidate confidence level ( If the threshold is within a range, such as 0.5 to 0.8, then the candidate will not be adopted and the protection strategy will be triggered or the system will fall back to a more conservative candidate. (4) If the candidate selection fails or is not feasible for (N) consecutive times (e.g., 1 to 3 times), the protection mode is entered directly.

[0074] The preferred rule for the final action selection is: select the candidate with the lowest risk and confidence level that meets the threshold from the feasible candidates; when there is conflict or uncertainty, give priority to more conservative strategy candidates.

[0075] S5. Perform a motion feasibility check on the final action; if the feasibility check fails or the confidence level is lower than the threshold, revert to the action candidate or trigger protection strategy under a lower risk strategy. The motion feasibility check is based at least on one or more constraints, including dynamic boundary, flight mode executability, fence boundary and obstacle distance, to determine whether the final action or skill atomic parameters are executable; when it is determined that it is not executable, it falls back to the action candidate under a lower risk strategy or triggers a protection strategy.

[0076] Before issuing the final action, this invention performs the final action. Or its parameter set, perform a motion feasibility check. The feasibility check includes at least one or more of the following constraints: (1) Dynamic boundary constraints: velocity, acceleration, and angular velocity must not exceed limits; (2) Flight mode executable constraints: Whether such commands are allowed in the current flight mode of Ardupilot (e.g., GUIDED allows speed / heading control, LOITER is used for hovering, and RTL / LAND is for safety actions). (3) Fence and obstacle departure constraints: Do not cross the boundary, do not approach the fence boundary, and do not be less than the minimum obstacle departure distance threshold; (4) Link and status health constraints: High-risk actions are prohibited when the communication link is abnormal or the flight control status is abnormal.

[0077] If the feasibility check fails, a rollback is executed: preferably rollback to a candidate action under a lower risk strategy, or directly trigger the protection strategy to enter a safe action such as hovering / returning to home / landing.

[0078] Existing technologies typically output a single action and lack multi-level strategies. This invention outputs safe, compromise, and aggressive multi-level candidate actions from a large model. The risk assessment and hierarchical decision-making module then automatically selects the appropriate action based on constraints such as maximum speed, minimum obstacle distance, fences, and low battery. When an action is not feasible or the confidence level is insufficient, the system backs down to a more conservative strategy or triggers protection. This ensures that risks are controllable and safety is prioritized from a system mechanism perspective.

[0079] S6. Map the final action to one or more predefined parameterized motion skill atoms, and manage the entry conditions, execution process and exit conditions of the motion skill atoms by a state machine; The motion skill atom includes at least one or more of the following: takeoff, landing, hovering, straight movement, backward movement, left movement, right movement, yaw rotation, ascent, descent, and return. Each skill atom contains one or more adjustable parameters, including speed, angular velocity, duration, target altitude, or turning angle. Each skill atom is a trajectory segment that a UAV can perform locally autonomous control.

[0080] The final action ( ) is mapped to one or more predefined parameterized motion skill atoms ( The entry, execution, exit, and exception transition of skill atoms are managed by a state machine. A motor skill atom includes at least (example, not a limitation): a) Hovering for t seconds; b) Travel straight along the current heading at speed v for t seconds; c) Travel the distance L in a straight line along the current heading at speed v; d) or( ): with angular velocity ( Rotate for t seconds or rotate ( )horn; e) , Vertical speed control; f)...... g) Return h) :landing.

[0081] Each of the aforementioned motion skill atoms is an independent motion trajectory segment, which can be defined by the length or time of the trajectory segment. Each motion skill atom is an independent, autonomously executable trajectory segment that can provide closed-loop feedback based on the UAV's own state perception system, ensuring the safe execution of each motion skill atom.

[0082] For example, The drone flies forward 1 meter at a speed of 0.5 m / s, using its own positioning system (such as Simultaneous Localization and Mapping (SLAM) or UWB providing current drone location information) to provide position feedback. The movement skill is completed atomically when the drone has traveled 1 meter. The distance traveled straight along the current heading is L; obstacle recognition system, positioning system, etc.

[0083] Please see Figure 2 The state machine includes at least the states of Idle, Skill Execution, Safety Protection, and Landing, and supports transition from any state to Safety Protection under abnormal conditions; Idle: waiting for valid operation; Skill Execution: the skill primitive is running, and the BCI input is locked or suppressed; Safety Protection: input when a risk or abnormal situation is detected; Landing: execute landing, and finally terminate safely.

[0084] During skill execution, to avoid accidental triggering and command conflicts, this invention performs locking, suppression, or delay processing on brain-controlled input: Lock: Ignore newly entered mind control event commands; Suppression: Only allow high-priority security events (such as HOVER / LAND / RTL) to interrupt execution; Delay: Cache new events until the current skill atom is completed before participating in the next round of intent reconstruction.

[0085] Preferably, the unlocking condition for locking / suppressing / delaying is triggered by the state machine based on the completion state of the skill atom, the release of the abnormal state, or a user confirmation event.

[0086] Existing technologies typically modify actions upon receiving commands, which can easily lead to concurrent conflicts when brain-controlled mis-triggers are frequent. This invention maps the final action to parameterized skill atoms, with a state machine managing entry / execution / exit. During execution, a locking / suppression / delay mechanism is employed, and execution is only interrupted by high-priority safety events (hovering / returning / landing). This significantly reduces attitude and trajectory jitter caused by command conflicts, thereby improving execution reliability.

[0087] S7. The motion skill atoms are executed in a closed loop through the UAV flight control and flight control interface, and safety boundary monitoring is performed during the execution process; when the preset protection trigger conditions are met, one or more protection strategies are executed, such as hovering, slow descent, return to home or landing. During the execution of motor skill atoms, the input of mind-controlled event commands is locked, suppressed, or delayed to avoid accidental triggering or command conflicts.

[0088] Preferably, the suppression process only allows high-priority security event instructions to interrupt the execution of the current skill atom, and the high-priority security events include at least one or more of hovering, returning to home, and / or landing.

[0089] The protection triggering conditions include at least: confidence level below a threshold, obstacle distance less than a threshold, triggering fence boundary, low battery threshold triggering, communication abnormality or flight status abnormality; and the protection triggering conditions also include failure to parse or verify the output of the large model.

[0090] Control commands are sent to the flight controller via the flight control interface, and the motion skill atoms are executed in a closed loop. The control commands include at least one or more of the following types: (1) Speed / heading control commands: used to realize skills such as straight-line movement, rotation, and ascent / descend; (2) Mode switching commands: used to enter safe modes such as hover hold (e.g., LOITER or equivalent hold mode), return to home (RTL), and landing (LAND); (3) Task / Action Commands: Used for advanced actions such as takeoff, landing, and return.

[0091] The system detects link health through a heartbeat / acknowledgment / timeout mechanism: when heartbeat is lost, acknowledgment times out, or the flight control status is abnormal, high-risk candidate actions are prohibited and protection strategies are triggered.

[0092] When the protection trigger conditions are met (including but not limited to: large model candidate confidence is below the threshold, motion feasibility check fails, minimum obstacle distance is less than the threshold, fence triggering or approach, low battery threshold triggering, communication abnormality or flight status abnormality), the present invention performs a safety fallback: preferably immediately enters hover hold (LOITER or GUIDED zero speed), and performs slow landing (LAND) or return to home (RTL) if the abnormality continues, to ensure the stability and safety of the drone.

[0093] Existing technologies often rely on operator experience or fragmented safety logic. This invention integrates insufficient confidence, insufficient obstacle clearance, fence triggering, low battery, link anomaly, abnormal flight status, and parsing verification failure into unified protection triggering conditions, and explicitly executes strategies such as hovering, slow descent, return to home, and landing. This enables the UAV to maintain a safe state under uncertain input or abnormal conditions, reduces the risk of crashes and debugging costs, and improves mission completion efficiency.

[0094] Example 1 A large-model-based brain-controlled drone resilience control system is provided. This system can be used to implement the aforementioned large-model-based brain-controlled drone resilience control method. Specifically, the large-model-based brain-controlled drone resilience control system includes an identification module, an encoding module, a reconstruction module, a decision-making module, an inspection module, a management module, and a monitoring module.

[0095] The identification module is used to collect and identify the user's brain signals and output discrete brain control event commands. The discrete brain control event commands are accompanied by at least identification confidence and / or signal quality indicators. The encoding module is used to serialize and encode the current discrete mind control event command and the historical discrete mind control event command within a sliding window of a preset time length, forming a command sequence, and then encode the command sequence into a symbol sequence that can be processed by a large model. The reconstruction module is used to receive the symbol sequence and the UAV state, and output at least one action candidate and the corresponding candidate confidence level; The decision module is used to perform risk assessment and hierarchical decision-making on the action candidates, select the final action based on the confidence level of the candidates and the risk assessment results, and parse and verify the structured fields output by the large model before making the decision. The inspection module is used to perform a motion feasibility check on the final action. When the feasibility check fails or the candidate confidence level is lower than the threshold, a rollback mechanism or protection strategy is triggered. The management module is used to map the final action that passes the feasibility check to one or more predefined parameterized motion skill atoms, and the state machine manages the entry conditions, execution process and exit conditions of the motion skill atoms. The monitoring module is used to execute the motion skill atoms through the UAV flight control and closed loop via the flight control interface. During the execution, it performs safety boundary monitoring and executes one or more protection strategies such as hovering, slow descent, return to home, or landing when the preset protection trigger conditions are met.

[0096] This invention provides a terminal device comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, graphics processing units (GPUs), tensor processing units (TPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve corresponding method flows or corresponding functions. The processor described in this embodiment can be used for the operation of a brain-controlled unmanned aerial vehicle (UAV) resilience control method based on a large model, including: The system collects and identifies the user's EEG signals to obtain discrete mind-controlled event commands. These commands include at least an identification confidence level and / or signal quality index. Within a preset time window, the current and historical discrete mind-controlled event commands are serialized and encoded to form a command sequence. This sequence is then encoded into a symbolic sequence that can be processed by a large model. The symbolic sequence and the drone's status are input into a large model intent reconstruction module, which outputs at least one action candidate and its corresponding confidence level. Risk assessment and hierarchical decision-making are performed on the action candidates, based on their confidence levels and risk assessment results. The system evaluates the results and selects the final action; it performs a motion feasibility check on the final action; if the feasibility check fails or the candidate confidence level is below a threshold, it reverts to a lower-risk action candidate or triggers a protection strategy; it maps the final action that passes the feasibility check to one or more predefined parameterized motion skill atoms, and the state machine manages the entry conditions, execution process, and exit conditions of the motion skill atoms; it executes the motion skill atoms through the UAV flight control and in a closed loop via the flight control interface, and performs safety boundary monitoring during the execution process; when the preset protection trigger conditions are met, it executes one or more protection strategies, such as hovering, slow descent, return to home, or landing.

[0097] Please see Figure 4The terminal device is a computer device. In this embodiment, the computer device 60 includes a processor 61, a memory 62, and a computer program 63 stored in the memory 62 and executable on the processor 61. When executed by the processor 61, the computer program 63 implements the large-model-based brain-controlled drone resilience control method described in this embodiment. To avoid repetition, details are omitted here. Alternatively, when executed by the processor 61, the computer program 63 implements the functions of each model / unit in the large-model-based brain-controlled drone resilience control system described in this embodiment. To avoid repetition, details are omitted here.

[0098] Computer device 60 can be a desktop computer, laptop, handheld computer, cloud server, or other computing device. Computer device 60 may include, but is not limited to, a processor 61 and a memory 62. Those skilled in the art will understand that... Figure 4 This is merely an example of computer device 60 and does not constitute a limitation on computer device 60. It may include more or fewer components than shown, or combine certain components, or different components. For example, computer device may also include input / output devices, network access devices, buses, etc.

[0099] The processor 61 may be a Central Processing Unit (CPU), or other general-purpose processors, graphics processing units (GPUs), tensor processing units (TPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0100] The memory 62 can be an internal storage unit of the computer device 60, such as a hard disk or memory of the computer device 60. The memory 62 can also be an external storage device of the computer device 60, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on the computer device 60.

[0101] Furthermore, the memory 62 may include both internal storage units of the computer device 60 and external storage devices. The memory 62 is used to store computer programs and other programs and data required by the computer device. The memory 62 can also be used to temporarily store data that has been output or will be output.

[0102] Please see Figure 5 The terminal device is an electronic device 600, which is manifested in the form of a general-purpose computing device. The components of the electronic device may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different platform components (including storage unit 620 and processing unit 610), a display unit 640, etc.

[0103] The storage unit stores program code, which can be executed by the processing unit 610 to perform the steps described in the method section of this specification according to various exemplary embodiments of the present invention. For example, the processing unit 610 can perform actions such as... Figure 1 The steps are shown in the figure.

[0104] Storage unit 620 may include a readable medium in the form of a volatile storage unit, such as random access memory (RAM) 6201 and / or cache memory 6202, and may further include a read-only memory (ROM) 6203.

[0105] Storage unit 620 may also include a program / utility 6204 having a set (at least one) program module 6205, such program module 6205 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0106] Bus 630 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the multiple bus structures.

[0107] Electronic device 600 can also communicate with one or more external devices 700 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 600, and / or with any device that enables electronic device 600 to communicate with one or more other computing devices (e.g., router, modem). This communication can be performed via input / output interface 650. Furthermore, electronic device 600 can also communicate with one or more networks (e.g., local area network, wide area network, and / or public network, such as the Internet) via network adapter 660. Network adapter 660 can communicate with other modules of electronic device 600 via bus 630. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage platforms.

[0108] Example 2 This invention also provides a storage medium, specifically a computer-readable storage medium, which is a memory device in a terminal device for storing programs and data. It is understood that the computer-readable storage medium here can include both built-in storage media in the terminal device and extended storage media supported by the terminal device; it can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which can be one or more computer programs (including program code). More specific examples of the computer-readable storage medium include: an electrical connection with one or more wires, a portable disk, a hard disk, random access memory, read-only memory, erasable programmable read-only memory, optical fiber, portable compact disk read-only memory, optical storage device, magnetic storage device, or any suitable combination thereof.

[0109] Computer-readable storage media also include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium can also be any readable medium other than a readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium can be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, radio frequency, etc., or any suitable combination thereof.

[0110] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0111] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the brain-controlled unmanned aerial vehicle resilience control method based on a large model in the above embodiments; one or more instructions in the computer-readable storage medium are loaded and executed by the processor in the following steps: The system collects and identifies the user's EEG signals to obtain discrete mind-controlled event commands. These commands include at least an identification confidence level and / or signal quality index. Within a preset time window, the current and historical discrete mind-controlled event commands are serialized and encoded to form a command sequence. This sequence is then encoded into a symbolic sequence that can be processed by a large model. The symbolic sequence and the drone's status are input into a large model intent reconstruction module, which outputs at least one action candidate and its corresponding confidence level. Risk assessment and hierarchical decision-making are performed on the action candidates, based on their confidence levels and risk assessment results. The system evaluates the results and selects the final action; it performs a motion feasibility check on the final action; if the feasibility check fails or the candidate confidence level is below a threshold, it reverts to a lower-risk action candidate or triggers a protection strategy; it maps the final action that passes the feasibility check to one or more predefined parameterized motion skill atoms, and the state machine manages the entry conditions, execution process, and exit conditions of the motion skill atoms; it executes the motion skill atoms through the UAV flight control and in a closed loop via the flight control interface, and performs safety boundary monitoring during the execution process; when the preset protection trigger conditions are met, it executes one or more protection strategies, such as hovering, slow descent, return to home, or landing.

[0112] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0113] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0114] Example 3 System module block diagram and engineering interface 1) Brain-controlled signal acquisition and recognition module Using SSVEP / P300 as an example of brain-controlled signal implementation, SSVEP implementation: Spectrum / CCA identification is performed on multi-frequency stimulus targets to output discrete intent categories (such as "forward / backward / left / right / hover / takeoff / landing") and confidence levels p.

[0115] P300 Implementation: ERP detection outputs the target selection event and confidence level p under the Oddball stimulus paradigm.

[0116] Example of output structure (event object): type: {TAKEOFF, LAND, HOVER, FORWARD, BACKWARD, LEFT, RIGHT, YAW_L, YAW_R, UP, DOWN} timestamp: t conf:

[0117] quality: q (such as channel availability, artifact rate, etc.) 2) Serialization encoding module (sliding window) Set the window length T (e.g., an optional range of 1–5 s, which can be adaptively adjusted according to signal quality), sort the events within the window by time to form a sequence, and encode it as a token: A token can include: event type, event interval Δt bucket, confidence level bucket, quality level, etc.

[0118] Structured input, such as JSON or a sequence of key-value pairs, can also be used as an alternative implementation.

[0119] 3) Large Model Intent Reconstruction Module (Resilience Control Center) Input: Window token sequence + drone state S (altitude, speed, mode, battery level, fence status, obstacle distance, etc., not limited). Output: Set of action candidates and their confidence scores, for example: Safe candidates: {HOVER(2s), SLOW_FORWARD(v=0.3m / s, t=1s)} Compromise candidate: {FORWARD(v=0.8m / s, t=1.5s)} Radical candidate: {FORWARD(v=1.5m / s, t=2s), YAW(ω=30° / s, t=1s)} Note: This does not limit the specific architecture and scale of the large model, but aims to achieve the functional features of sequence reasoning and intent completion.

[0120] 4) Risk assessment and tiered decision-making module (incorporating real-world security constraints) Set at least the following security constraints: Maximum speed constraint: (Can be configured according to mode / environment, such as lower indoors and higher outdoors); Minimum obstacle distance constraint: (Obtained from ranging / vision / LiDAR / ultrasound); Fence constraints: Geofencing or virtual fencing, including horizontal and vertical boundaries; Low battery threshold: It will force the system to enter a return-to-home / landing strategy.

[0121] The risk assessment outputs a risk score R and a risk level L. The decision rule can be: like (For example, 0.6 can be used as a range rather than a fixed value) → directly enter the protection strategy (LOITER / HOVER + optional slow descent / RTL); like Or trigger the fence boundary → retreat to the safety policy candidate or perform brake / hover; like → Force RTL or LAND, and do not perform aggressive actions.

[0122] 5) Exercise Feasibility Check Module Perform dynamic and model feasibility verification on the parameters (v, ω, t, h, etc.) of the candidate actions: Velocity / acceleration / angular velocity must not exceed limits; Does the flight control system allow such commands in the current mode (e.g., speed setting is allowed in GUIDED mode, hovering mode is suitable for hovering, and RTL / LAND return-to-home / landing mode is a safe action)?

[0123] If it is not feasible, roll back or trigger protection.

[0124] 6) Motor skill atomic mapping module + state machine management module Define a set of skill atoms (example): g_hover(t): Hover for t seconds (corresponding to the expected speed of the flight controller being 0); g_forward(v, t): Continues forward along the body at a velocity v for t. g_yaw(ω, t) or g_yaw(θ): yaw angular velocity ω for duration t or rotation angle θ; g_up(vz, t) / g_down(vz, t): Vertical speed control; g_rtl: return home; g_land: land.

[0125] Example state machine states: Idle, Skill Execution, Safety Protection, Landing. During skill atom execution, mind-controlled input is: locked / suppressed / delayed.

[0126] 7) Flight control command generation and transmission module As an example, if the flight controller is Ardupilot and the flight controller interface protocol is MAVLink, then the system sends control commands to Ardupilot through MAVLink, including at least one or more of the following implementations: g_forward, g_yaw, and g_up are implemented through speed control messages (such as in the form of speed vectors / yaw rates); g_hover (LOITER), g_rtl (RTL), and g_land (LAND) are implemented through mode switching and task commands. The health of the link is detected through a heartbeat and confirmation mechanism, and SafetyProtect is triggered when the link is abnormal.

[0127] Example 4 1. Collect EEG data and identify the event command u(t) based on SSVEP and / or P300, and output the confidence level p(t) and quality q(t); 2. Construct an event sequence within the sliding window W and encode it as a token sequence X; 3. Input X and the drone's state S into a large model, and output a set of multiple risk candidate actions. and confidence level; 4. Calculate risk assessment R and select strategies based on maximum speed, minimum obstacle distance, fence, and low battery threshold; 5. Feasibility check: If not feasible, fall back to a more conservative candidate; if conf < If continuous infeasibility fails, then a fallback protection mechanism will be implemented; 6. Map the final action to a sequence of skill atoms, managed by a state machine, and lock / suppress the input; 7. Send commands to the flight controller (such as Ardupilot) via the flight controller interface (such as MAVLink) for closed-loop execution and continuous safety monitoring; if triggered, enter protection modes such as LOITER / RTL / LAND.

[0128] Example 5 The protection policy is triggered when any of the following conditions are met: 1) The confidence level of the large model is below the threshold; 2) The distance to the obstacle is less than the minimum threshold; 3) Triggering the fence boundary or approaching the fence; 4) Low battery threshold trigger; 5) MAVLink communication error or flight control status error.

[0129] Protection policy execution order (configurable as an example): 1) Immediately enter hover hold (LOITER or GUIDED zero speed); 2) If recovery does not continue, descend slowly and prepare for landing (LAND). 3) If conditions for returning to home are met and battery power allows, then execute RTL; 4) Perform a forced landing / intermittent descent if necessary.

[0130] Example 5: Degradation Mode Example When the large model is unavailable or the communication link is poor, the system enters degradation mode to ensure that the system can still operate safely under the worst-case scenario. Replace large model intent reconstruction with a rule engine / finite state table; Only safe skill atoms are allowed (hovering, low-speed straight travel, stop, return, landing); Increase confidence threshold and risk level, reduce and increase ; Example 6: Token Encoding Example within a Sliding Window A sliding window-based event serialization and tagging encoding process is used to convert discrete brain control events into sequences that the model can process.

[0131] Multiple brain-computer interface events within the window are converted into tokens, with each event contributing multiple semantic tokens. The tokens are arranged chronologically to form an input sequence for use by a large language model.

[0132] The mathematical definition of a discrete mind control event is as follows: the (i)th mind control event is defined as:

[0133] in: Event categories (FORWARD, LEFT, HOVER, ...); For event timestamps; ) is used to identify confidence level; This is a signal quality indicator.

[0134] Define a sliding window event sequence within the time window. Within, the event sequence is as follows:

[0135] Define the token mapping function as: (1) Event type Token

[0136] For example

[0137] (2) Time interval Token (Δt bucket)

[0138] Define the interval discretization function:

[0139] For example:

[0140] (3) Confidence Token

[0141]

[0142] (4) Quality Token

[0143] in, This is a quality level mapping function.

[0144] For a single event i The corresponding token subsequence is:

[0145] A complete token sequence (LLM input)

[0146] The sequence is used as input to the large model for intent reconstruction and continuous parameter generation.

[0147] Discrete labeled sequences encode semantic intent, time, confidence, and signal reliability, enabling large language models to infer continuous motion parameters under structured constraints, rather than directly regressing control commands from noisy brain signals.

[0148] Example 7 To verify the feasibility and beneficial effects of this invention, it was validated in the Ardupilot SITL simulation environment and the brain control command offline playback environment.

[0149] The system control chain is as follows: Discretized brain control events (including confidence / quality) → Sliding window serialization → Large model intent reconstruction (structured input / output) → Parsing verification → Risk stratification selection / rollback → Motion feasibility check → Skill atom / state machine execution → MAVLink distribution → Ardupilot execution → Safety monitoring and protection mode (LOITER / RTL / LAND).

[0150] The comparison object is set as follows: Baseline A (existing technology): Discrete brain control commands are directly mapped to flight control quantities / modes (optional simple majority voting), without large-scale model intent reconstruction, risk stratification, structured parsing verification, or skill atom locking mechanism; Option B (the present invention): adopts the complete link of the present invention (intent reconstruction + hierarchical decision-making + feasibility + skill atoms + protection strategy).

[0151] To simulate the non-ideal case of actual SSVEP / P300 decoding, the following is introduced into the discrete event sequence: False trigger rate (randomly inserted error events); Omission rate (random missing events); Latency jitter (random delay of event timestamps); Confidence fluctuations and quality degradation (conf / quality drift over time).

[0152] Under the aforementioned perturbation, baseline A and invention B are compared under the same task.

[0153] Evaluation indicators: Mission success rate: The percentage of missions that reach the target location, successfully navigate around obstacles, or do not trigger loss of control. Minimum obstacle clearance distance: the minimum distance during flight. ; Protection trigger count and effectiveness: After protection is triggered, does it stably enter LOITER / RTL / LAND, and does it avoid collision / boundary crossing?

[0154] Typical simulation results: Under the same perturbation injection conditions, the present invention exhibits higher stability and security compared to baseline A. Typical results are shown in Table 1.

[0155] Table 1. Typical simulation comparison results (baseline A versus invention B)

[0156] The results show that, under conditions of discrete brain control commands, noise, and time delay fluctuations, this invention, through large model intent reconstruction, risk layering and fallback, motion feasibility verification, and execution of motion skill atomic state machine with safety fallback strategy, can significantly reduce the proportion of control jitter and conflicting commands, improve task success rate and safety margin, and achieve verifiable safety mode implementation under conditions such as fence / low power / link anomaly, proving that this invention has good engineering feasibility and beneficial effects.

[0157] The method of the present invention has the following characteristics: 1) Resilient hub for discrete brain control event serialization and large model intent reconstruction: The brain control recognition output is upgraded from a single command to an event sequence within a sliding window (including event type, timestamp / interval, recognition confidence, signal quality indicators, etc.), and input into a large model for intent reconstruction / completeness / error correction, outputting more stable and executable control intent or action candidates, rather than directly executing a single classification result.

[0158] 2) Large-scale model engineering implementation mechanism with structured input / output constraints: The input of the large model is organized with structured fields, and the output uses structured candidate action fields, so that the output of the large model is parsable, verifiable and controllable, avoiding the ambiguity and uncontrollable risks brought by free text.

[0159] 3) Parsing and verification, parameter pruning, and a fallback mechanism for failures: Before the large model output enters the decision / flight control system, a parsing and verification module is set up (field integrity, type verification, parameter range pruning, and default filling); when parsing or verification fails, the output is not sent and a fallback or protection strategy is triggered. This is the key to how this invention introduces the large model into the control link, enabling it to meet or even enhance the system's resilience and security.

[0160] 4) Multiple risk level candidates (safe / compromise / aggressive) and risk stratification decision-making and backoff: The large model outputs a set of multiple strategy candidate actions (at least two levels, with three levels preferred: safe / compromise / aggressive), and performs risk assessment and selection in combination with UAV status and safety constraints; when the risk is high, the confidence level is insufficient or it is not feasible, it automatically backoffs to a more conservative strategy or directly enters the protection strategy, so as to achieve resilient control of safety first and efficiency later.

[0161] 5) Motion feasibility check (hard constraint closed loop): Perform a motion feasibility check on the final action and its parameters, including at least dynamic boundaries, flight mode executability, fence / obstacle clearance constraints, link / flight control health status, etc.; if the check fails, back off or protect; ensure that the candidate action is not only reasonable, but also can be executed in the current flight control mode.

[0162] 6) Skill Atom Parameterized Mapping and State Machine Closed-Loop Execution: The final action is mapped to parameterizable motion skill atoms (hovering, straight-line, rotation, ascent and descent, return to home, landing, etc.), and the state machine manages entry / execution / exit and abnormal transfer. It is stably executed through flight control closed loop, realizing the decoupling of brain control triggering and UAV autonomous closed-loop execution, reducing the dependence on continuous high-quality brain control signals.

[0163] 7) Input locking / suppression / delay and safety priority interruption mechanism during execution: During the atomic execution of motor skills, the brain-controlled input is locked / suppressed / delayed to prevent concurrent command conflicts; and it is stipulated that only high-priority safety events (hovering / returning / landing, etc.) are allowed to interrupt the execution; this is the key to combating brain-controlled false triggers and improving execution stability.

[0164] 8) Unified set of safety monitoring and protection strategy trigger conditions: Insufficient confidence, failed parsing verification, insufficient obstacle clearance, fence triggering / approach, low battery threshold, communication anomaly, abnormal flight status, etc. are uniformly included in the protection trigger conditions; after triggering, hovering / descent / return to home / landing and other safety fallback measures are implemented. This reflects a full-link safety closed loop from input to execution.

[0165] In summary, this invention provides a robust control method and system for brain-controlled drones based on a large model. Under conditions of discrete brain-controlled commands, high noise, latency fluctuations, and the coexistence of false triggers and omissions, it can obtain more stable executable intentions through sliding window serialization encoding and large model intention reconstruction. Simultaneously, through risk-layered candidate generation, risk assessment and rollback, motion feasibility checks, and skill atom / state machine management, it achieves safe constraints and stable execution of drone control commands. Furthermore, it promptly triggers protection strategies such as hovering, return-to-home, or landing in situations such as parsing verification failure, insufficient confidence, insufficient obstacle clearance, fence triggering, low battery, and link anomalies, thereby significantly improving the stability, safety, and robustness of brain-controlled drone control.

[0166] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0167] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0168] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0169] In the embodiments provided by this invention, it should be understood that the disclosed devices / terminals and methods can be implemented in other ways. For example, the device / terminal embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0170] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0171] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0172] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random-access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0173] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0174] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0175] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0176] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A resilient control method for brain-controlled unmanned aerial vehicles based on a large model, characterized in that, Includes the following steps: S1. Collect and identify the user's brainwave signals to obtain discrete brain control event instructions, wherein the discrete brain control event instructions are accompanied by at least identification confidence and / or signal quality indicators. S2. Within a sliding window of a preset time length, the current discrete mind control event command and the historical discrete mind control event command are serialized and encoded to form a command sequence, and the command sequence is encoded into a symbol sequence that can be processed by a large model. S3. Input the symbol sequence and the UAV state into the large model intent reconstruction module, and the large model intent reconstruction module outputs at least one action candidate and the corresponding candidate confidence. S4. Perform risk assessment and hierarchical decision-making on the candidate actions, and select the final action based on the confidence level of the candidates and the risk assessment results; S5. Perform a motion feasibility check on the final action; if the feasibility check fails or the candidate confidence level is lower than the threshold, revert to the action candidate under a lower risk strategy or trigger a protection strategy. S6. The final action that passes the feasibility check is mapped to one or more predefined parameterized motion skill atoms, and the state machine manages the entry conditions, execution process and exit conditions of the motion skill atoms. S7. The motion skill atoms are executed in a closed loop through the UAV flight control and the flight control interface, and safety boundary monitoring is performed during the execution process; When the preset protection trigger conditions are met, one or more protection strategies are executed, such as hovering, slow descent, return to home, or landing.

2. The resilient control method for brain-controlled unmanned aerial vehicles based on a large model according to claim 1, characterized in that, In step S1, the EEG signal is preprocessed after acquisition. The preprocessing includes at least one or more of bandpass filtering, power frequency suppression, artifact suppression, and baseline correction. The discretized brain-controlled event commands are identified based on the SSVEP paradigm and / or P300 paradigm. The event types of the discretized brain-controlled event commands include at least one or more of takeoff, landing, hovering, forward, backward, leftward, rightward, leftward, rightward, ascent, descent, and return.

3. The resilient control method for brain-controlled unmanned aerial vehicles based on a large model according to claim 1, characterized in that, In step S2, the length of the sliding window can be adaptively adjusted according to the signal quality index. When the signal quality decreases or the drift increases, the window length is increased; when the signal quality improves and the response requirement is higher, the window length is decreased. The symbol sequence is organized in the form of structured fields, including at least one or more of the following: event sequence field, UAV status field, and safety constraint field; each discrete mind control event instruction in the instruction sequence includes at least two or more of the following: event type, timestamp, recognition confidence level, and signal quality index.

4. The resilient control method for brain-controlled unmanned aerial vehicles based on a large model according to claim 1, characterized in that, In step S3, the drone status includes at least one or more of the following: flight mode, altitude, speed vector, battery level, fence status, minimum obstacle distance, and communication link status; the action candidates are output in the form of structured fields, including at least one or more of the following: risk level field, action type field, parameter field, and candidate confidence field; the action candidates include at least a candidate set corresponding to two or three levels of risk strategies among safe candidates, compromise candidates, and aggressive candidates.

5. The resilient control method for brain-controlled unmanned aerial vehicles based on a large model according to claim 1, characterized in that, In step S4, before performing risk assessment and hierarchical decision-making, the structured fields output by the large model intent reconstruction module are parsed and verified. The parsing and verification include at least field integrity checks, parameter range pruning, and / or default filling. When parsing or verification fails, the action candidate is not issued and the protection strategy is triggered or the action candidate is rolled back to a lower risk strategy. The risk assessment is based on at least two or more of the following constraints: maximum speed constraint, minimum obstacle distance constraint, geofence constraint, and low battery threshold constraint.

6. The resilient control method for brain-controlled unmanned aerial vehicles based on a large model according to claim 1, characterized in that, In step S5, the motion feasibility check is based on at least one or more of the following: dynamic boundary constraints, flight mode executability constraints, fence boundary constraints, and obstacle distance constraints, to determine whether the final action or skill atomic parameters are executable; the protection strategy includes at least one or more of the following: entering hovering mode, return-to-home mode, and / or landing mode.

7. The resilient control method for brain-controlled unmanned aerial vehicles based on a large model according to claim 1, characterized in that, In step S6, the motion skill atom includes at least one or more of the following: takeoff, landing, hovering, straight movement, backward movement, left movement, right movement, yaw rotation, ascent, descent, and return. Each skill atom contains one or more adjustable parameters, including speed, angular velocity, duration, target altitude, or turning angle. The state machine includes at least an idle state, a skill execution state, a safety protection state, and a landing state, and supports transitioning from any state to the safety protection state under abnormal conditions.

8. The resilient control method for brain-controlled unmanned aerial vehicles based on a large model according to claim 7, characterized in that, During the skill execution state, new discretized mind control event commands are locked, suppressed, or delayed. The locking process ignores the newly entered discretized mind control event commands, the suppression process only allows high-priority safety event commands to interrupt the execution of the current skill atom, and the delay process caches the new discretized mind control event commands until the current skill atom is completed before participating in the next round of intent reconstruction. The high-priority safety events include at least one or more of hovering, returning to home, and / or landing.

9. The resilient control method for brain-controlled unmanned aerial vehicles based on a large model according to claim 1, characterized in that, In step S7, the protection triggering conditions include at least: candidate confidence level below a threshold, obstacle distance less than the minimum obstacle distance threshold, triggering fence boundary, low battery threshold triggering, communication anomaly, flight status anomaly, and large model output parsing or verification failure; the control commands issued by the flight control interface include at least one or more of the following: speed / heading control commands, mode switching commands, and task / action commands; when the large model is unavailable or the communication link is poor, the system enters a degraded mode, using a rule engine / finite state table to replace the large model intent reconstruction, allowing only safety skills to be executed atomically and increasing the confidence threshold and risk level.

10. A resilient control system for brain-controlled unmanned aerial vehicles based on a large model, characterized in that, include: The recognition module is used to collect and recognize the user's brain signals and output discrete brain control event commands, wherein the discrete brain control event commands are accompanied by at least recognition confidence and / or signal quality indicators. The encoding module is used to serialize and encode the current discrete mind control event command and the historical discrete mind control event command within a sliding window of a preset time length, forming a command sequence, and then encode the command sequence into a symbol sequence that can be processed by a large model. The reconstruction module is used to receive the symbol sequence and the UAV state, and output at least one action candidate and the corresponding candidate confidence level; The decision module is used to perform risk assessment and hierarchical decision-making on the action candidates, select the final action based on the confidence level of the candidates and the risk assessment results, and parse and verify the structured fields output by the large model before making the decision. The inspection module is used to perform a motion feasibility check on the final action. When the feasibility check fails or the candidate confidence level is lower than the threshold, a rollback mechanism or protection strategy is triggered. The management module is used to map the final action that passes the feasibility check to one or more predefined parameterized motion skill atoms, and the state machine manages the entry conditions, execution process and exit conditions of the motion skill atoms. The monitoring module is used to execute the motion skill atoms through the UAV flight control and closed loop via the flight control interface. During the execution, it performs safety boundary monitoring and executes one or more protection strategies such as hovering, slow descent, return to home, or landing when the preset protection trigger conditions are met.

Images