Unmanned aerial vehicle voice control full-link real-time security verification system

Abstract

This invention discloses a real-time security verification system for the entire voice control chain of unmanned aerial vehicles (UAVs), relating to the fields of intelligent UAV control and voice security technology. This invention constructs a high-risk homophone feature library for UAVs with risk weights, employs a lightweight edge-side model to perform secondary verification and error correction on ASR (Automatic Speech Recognition) results, and avoids the risk of misidentification of opposing commands from the source. It achieves accurate flight status perception through multi-source sensor data time synchronization, adaptive fusion filtering, and standardized finite state machines. Based on a hierarchical safety rule expert system, it performs parallel three-dimensional conflict detection of timing, state-command, and multiple command superposition, outputting hierarchical execution decisions. Legitimate commands are issued and executed, while high-risk commands are intercepted and trigger safety fallback actions. Simultaneously, execution results are collected and fed back to update the rule library, forming a closed-loop control system across the entire chain. This invention effectively reduces the misjudgment rate and security risks of voice control, is compatible with mainstream ASR engines and multi-brand UAV flight controllers, and is adaptable to offline edge-side deployment on consumer and industrial-grade UAVs, providing highly reliable security for UAV voice control.

Description

Technical Field

[0001] This invention relates to the field of intelligent control and voice interaction technology for unmanned aerial vehicles (UAVs), specifically to a real-time security verification system for the entire UAV voice control chain. Background Technology

[0002] With the large-scale application of drone technology in fields such as aerial photography, power line inspection, agricultural plant protection, and emergency rescue, voice control, as a human-computer interaction method that can free up hands and improve the efficiency of operation in complex scenarios, has become an important research and development direction in the field of drone human-computer interaction.

[0003] Most current mainstream drone voice control solutions directly use general-purpose automatic speech recognition (ASR) engines to recognize commands and then send the recognition results directly to the flight controller for execution. This type of solution suffers from the following insurmountable technical drawbacks in practical applications: The general-purpose ASR engine is optimized for everyday general dialogue scenarios, but it has not been specifically optimized for high-frequency homophones with completely contradictory meanings in drone control scenarios that could directly cause flight safety accidents after execution, nor has it differentiated control commands by risk level. In outdoor flight scenarios with strong noise environments such as wind noise and propeller noise, it is very easy to misidentify homophones such as "take off / land" and "ascend / descend," which poses a serious safety hazard from the source of command recognition, and the existing technology has no targeted source protection mechanism.

[0004] Existing security protection solutions have a single verification dimension. Most of them only perform basic threshold filtering on the confidence of ASR recognition results, or simply match a single command with the current flight status of the drone. They have not formed a closed-loop protection system covering the entire chain from command recognition to flight execution. They cannot identify complex and dangerous scenarios such as timing conflicts, state-command mismatches, and multiple command superposition conflicts, resulting in serious problems of blurred security boundaries.

[0005] Existing solutions often directly use the raw, single data output by the flight controller to determine the flight status without performing fusion filtering on multi-source sensor data. This fails to eliminate parameter errors caused by sensor noise, sampling delay, and data drift, resulting in insufficient accuracy in flight status determination. This can easily lead to misjudgments and omissions in command validity verification, and cannot provide a reliable benchmark for safety verification.

[0006] Existing solutions are mostly custom-developed for specific aircraft models and specific ASR engines, lacking standardized upstream and downstream adaptation interfaces. They are incompatible with mainstream commercial ASR engines and multi-brand drone flight control SDKs, and cannot adapt to refined safety rules based on aircraft model, operation scenario, and task type. Their generalization ability is weak, and large-scale deployment is difficult.

[0007] Among the published patent documents, CN120412564A discloses a method for parsing drone voice commands, which only implements general semantic confidence scoring and basic security interception, without setting up a specific error correction mechanism for high-risk homophones in drones; CN118915544A discloses a voice command error correction method, which only performs error correction based on historical dialogues for general scenarios, without risk classification and targeted optimization for drone control scenarios; CN202510928233A discloses a drone safety control method, which only performs safety potential field judgment based on simple environmental data, without constructing a flight state machine with multi-source data fusion and a multi-dimensional conflict reasoning mechanism. In summary, none of the existing technologies disclose a complete real-time security verification scheme for the entire process of drone voice control, and cannot fundamentally solve the common industry problems of high accidental touch rate, frequent command conflicts, and blurred security boundaries in drone voice control. Summary of the Invention

[0008] The purpose of this invention is to overcome the aforementioned deficiencies of the prior art and, in response to the specific security requirements of drone voice control scenarios, to provide a complete end-to-end real-time security verification solution, specifically addressing the following technical problems: To address the high risk of misidentification of high-risk homophones and the lack of source security protection in existing technologies for drones, and to achieve targeted verification and error correction of high-risk conflicting commands; To address the issues of insufficient accuracy in flight status perception and unreliable verification benchmarks in existing technologies, this paper proposes to achieve accurate identification of UAV flight status through multi-source data fusion, providing a stable and reliable benchmark for safety verification. To address the issues of existing technologies having a single security verification dimension and insufficient ability to identify complex conflict scenarios, a multi-dimensional parallel conflict detection mechanism is constructed to clarify the security boundaries of drone voice control. To address the issues of existing technologies lacking end-to-end closed-loop protection and having poor generalization adaptability, a full-process closed-loop verification system is constructed, from command recognition to flight execution, enabling refined rule adaptation and large-scale deployment across multiple scenarios.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A real-time security verification method for the entire voice control chain of drones includes the following core steps: S1 Source Identification and Verification: A high-risk homophone feature library for UAVs with risk weight classification is pre-constructed. The feature library contains high-risk homophone pairs / word sets that are semantically contradictory and may cause flight safety accidents after execution in UAV control scenarios, as well as the acoustic features, textual semantic features, and control context features of the corresponding commands. High-risk contradictory commands are set with the highest verification weight. The Top-N candidate recognition results, corresponding acoustic confidence scores, and contextual temporal command information output by the general ASR engine are obtained. If the recognition result falls into the high-risk homophone feature library, a lightweight speech classification model adapted to the edge deployment is used to perform secondary classification and confidence reordering of the candidate recognition results, and the verified control command with the highest confidence score and conforming to the control logic is output. If it does not fall into the library, the original recognition result is directly output as the verified control command. S2 Real-time Flight Status Perception: This system collects core flight control data and peripheral sensor data from the UAV in real time. The core flight control data includes altitude, three-axis velocity, three-axis attitude angles, GPS positioning data, and battery level and voltage. The peripheral sensor data includes obstacle avoidance radar / visual obstacle avoidance distance data and wind speed sensor data. Multi-source data with different sampling frequencies are time-stamped and synchronized. After filtering and preprocessing to remove impulse noise and high-frequency jump data, an adaptive fusion filtering algorithm adapted to the UAV's nonlinear motion model is used for fusion processing, outputting noise-free, low-latency core state parameters. A standardized finite state machine model is pre-constructed based on civil aviation UAV control specifications and the UAV's physical characteristics. This finite state machine model divides the UAV's entire lifecycle flight state into discrete states with clearly defined entry / exit conditions and parameter safety thresholds. Based on these core state parameters, the system determines and outputs the UAV's current precise flight state in real time. S3 Dynamic Command Legality Reasoning: A pre-built hierarchical safety rule expert system is constructed. This system includes basic safety rules, scenario-based safety rules, and conflict resolution rules, and defines fixed command priorities, from highest to lowest: emergency commands > safety commands > routine control commands > mission commands. Using the verified control commands output in step S1 and the real-time flight status output in step S2 as inputs, the system performs parallel conflict detection across three dimensions: temporal conflict detection, state-command conflict detection, and multi-command superposition conflict detection. Triggering any high-risk rule results in the command being deemed illegal. Based on the detection results, a tiered execution decision is output: legal commands are allowed execution; low-risk illegal commands are deferred with a warning; and high-risk illegal commands are intercepted with a safety fallback action trigger. S4 Full-link execution control and closed-loop feedback: Based on the execution decision output in step S3, legally granted instructions are converted into a format recognizable by the corresponding UAV flight control SDK and issued for execution; high-risk illegal instructions are directly intercepted and corresponding safety fallback actions are triggered; the execution results of flight control instructions are collected in real time and fed back to the high-risk homonym feature library in step S1 and the safety rule expert system in step S3, completing the incremental updates of the feature library and rule library, forming a full-link closed-loop control.

[0010] Furthermore, in step S1, the construction of the high-risk homonym feature library also includes collecting environmental noise samples, multi-accent speech samples, and historical misidentification samples of the corresponding ASR engine in the target application scenario, extracting the corresponding features and supplementing them to the feature library.

[0011] Furthermore, in step S2, the fusion processing logic of the adaptive fusion filtering algorithm is as follows: using high-frequency inertial data to compensate for the low-frequency delay of GPS positioning data, using altimeter data to supplement blind spot data for visual obstacle avoidance, and using absolute position data to correct the cumulative drift of inertial data.

[0012] Furthermore, in step S3, the safety fallback action includes at least one of hovering, returning to base, staying in place, and deceleration.

[0013] Furthermore, in step S1, the general ASR engine includes a commercial online ASR engine and an offline edge ASR engine.

[0014] Furthermore, in step S2, the peripheral sensor data also includes ambient light data, temperature and humidity data, and air pressure data.

[0015] Furthermore, in step S3: The timing conflict detection is to detect whether there is a timing logic conflict between the current instruction and the preceding instruction that has not been completed. The state-command conflict detection is used to detect whether the current command meets the entry / exit conditions and parameter safety thresholds corresponding to the current flight state. The multi-instruction superposition conflict detection is used to detect whether there is a risk of exceeding the flight safety boundary of the UAV after the current instruction is superimposed with multiple instructions executed in parallel.

[0016] Furthermore, in step S4, at least two mainstream commercial ASR engines are connected through a standardized upstream interface, and at least two mainstream brand drone flight control SDKs are connected through a standardized downstream interface.

[0017] Corresponding to the above method, the present invention also provides a real-time security verification system for the entire voice control chain of unmanned aerial vehicles, comprising: The source identification and verification module is used to pre-build a feature library of high-risk homophones for drones with risk weight classification, obtain the identification-related data output by the general ASR engine, complete the secondary verification and error correction of high-risk homophones, and output the verification control command. The flight status perception module is used to collect multi-source flight data of UAVs, complete data synchronization, preprocessing and fusion filtering, and judge and output the current flight status of UAVs in real time based on a pre-built standardized finite state machine model. The dynamic command legality reasoning module is used to pre-build a hierarchical safety rule expert system. It takes the verified control commands and real-time flight status as inputs, completes parallel conflict detection in three dimensions, and outputs hierarchical command execution decisions. The end-to-end execution control and closed-loop feedback module is used to issue or intercept commands based on execution decisions, while collecting flight control execution results and feeding them back to the corresponding modules to complete incremental updates of the feature library and rule library, thereby achieving end-to-end closed-loop control.

[0018] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the real-time security verification method for the entire link of UAV voice control as described above.

[0019] Compared with the prior art, the present invention has the following beneficial effects that are inevitable to be achieved by the technical solution, completely addresses the core defects in the background technology, and contains no false quantitative data that has not been experimentally verified: This invention constructs a phonetic near-word error correction mechanism specifically for UAVs, abandoning the conventional approach of indiscriminate optimization of general ASR. It performs targeted weighting and secondary verification on high-risk commands with semantic contradictions in UAV control scenarios, avoiding flight safety accidents caused by misrecognition of speech from the source of command recognition, and filling the gap in the existing technology for security protection at the source of UAV speech recognition.

[0020] This invention employs a flight state estimation algorithm based on multi-source data fusion to perform time synchronization, noise filtering, and adaptive fusion filtering on multi-source sensor data from UAVs. This effectively eliminates parameter errors caused by sensor noise, sampling delay, and data drift, significantly improving the accuracy of UAV flight state discrimination. It provides a stable and reliable benchmark for command legality verification and can greatly reduce the probability of misjudgment and missed judgment in security verification.

[0021] This invention constructs a three-dimensional parallel conflict detection mechanism and a hierarchical safety rule expert system, which breaks through the limitations of the single-dimensional verification of existing technologies. It can effectively identify complex and dangerous scenarios such as timing conflicts, state-instruction mismatches, and multiple instruction superposition conflicts that existing technologies cannot handle, clarify the safety boundaries of UAV voice control, and significantly reduce flight safety risks in voice control mode.

[0022] This invention constructs a serial verification closed loop of "source identification and secondary verification - precise flight status perception - dynamic legality reasoning - full-link execution control - closed-loop iterative optimization", forming a dual protection system for the entire process from voice input to flight control execution. At the same time, it realizes incremental updates of feature library and rule library through reverse feedback of execution results, which can continuously adapt to the security protection needs of different scenarios and has strong engineering applicability.

[0023] This invention constructs a standardized upstream and downstream adaptation interface that is compatible with mainstream commercial ASR engines, offline edge ASR engines, and multi-brand drone flight control SDKs. It can perform fine-grained safety rule adaptation according to different drone models, operating scenarios, and task types, and has strong generalization capabilities, which greatly reduces the threshold for large-scale application of drone voice control safety solutions.

[0024] The technical solution of this invention can be flexibly deployed on control terminal mobile devices and airborne edge computing modules, supports offline operation, and can be adapted to the full-scenario voice control needs of consumer-grade and industrial-grade drones, providing reliable security for the large-scale compliant application of drone voice control. Attached Figure Description

[0025] Figure 1 is an overall flowchart of the real-time security verification method for the entire link of UAV voice control described in this invention; Figure 2 is a flowchart of the UAV-specific phonetic similarity word correction algorithm described in this invention; Figure 3 is a flowchart of the flight state estimation algorithm based on multi-source data fusion described in this invention; Figure 4 is a flowchart of the dynamic instruction legality reasoning algorithm of the present invention; Figure 5 is an architecture block diagram of the real-time security verification system for the entire link of UAV voice control described in this invention; Detailed Implementation

[0026] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Those skilled in the art can fully implement the present invention, solve the corresponding technical problems, and obtain the beneficial effects of the present invention based on the core technical solutions disclosed in this specification without creative effort. Implementation

[0027] This embodiment discloses the core implementation logic of the present invention, which completely corresponds to the technical solution defined in the claims, as follows:

[0028] First, a high-risk homophone feature library for drones with risk weight classification is pre-constructed. The specific construction method is as follows: The complete set of UAV control commands was analyzed and categorized into three levels based on the safety risk level after execution: high risk, medium risk, and low risk. High-risk commands are those with completely contradictory meanings that will directly cause flight safety accidents after execution, including but not limited to homophonous word pairs / word sets such as "take-off / landing", "ascend / descend", "forward / backward", "fly left / fly right", "return home / cancel return home", and "power on / power off". Acoustic features, text semantic features, and control context features corresponding to the above high-risk instructions are collected. At the same time, environmental noise samples such as wind noise and blade noise in the target application scenario, multi-accent speech samples from different regions and age groups, and historical misidentification samples corresponding to the ASR engine are collected. The corresponding features are extracted and added to the feature library. Set risk weights for instructions in the feature library, with high-risk opposing instructions set to the highest verification weight to ensure priority processing during secondary verification.

[0029] After completing the feature library construction, obtain the Top-N candidate recognition results (N is a positive integer, which can be set by those skilled in the art according to the performance of the ASR engine, usually 3-5), the acoustic confidence of each candidate result, and the context timing instruction information that has been executed / to be executed in the preceding order; First, determine whether the recognition result falls into the high-risk homonymous word feature library. If it does not, directly output the original recognition result as the control command after verification. If it falls into the high-risk homonymous word feature library, start the lightweight speech classification model deployed on the adapted end side, perform secondary classification and confidence re-ranking on the Top-N candidate results, and finally output the correct command with the highest confidence and in line with the control context logic to complete the source recognition verification.

[0030] The lightweight speech classification model can be a conventional lightweight CNN model or a lightweight Transformer model adapted to mobile / airborne edge computing modules. It does not rely on cloud computing power and can run offline on the device. Those skilled in the art can select the model according to the computing power of the deployed device and complete the training and deployment of the model without creative labor.

[0031] First, the drone's flight control core data and peripheral sensor data are collected in real time through the drone's onboard communication link. The flight control core data includes altitude, three-axis speed, three-axis attitude angle, GPS positioning data, battery power and voltage. The peripheral sensor data includes obstacle avoidance radar / visual obstacle avoidance distance data and wind speed sensor data. Depending on the drone model configuration, ambient light data, temperature and humidity data, and air pressure data can also be collected. For multi-source sensor data with different sampling frequencies, data synchronization is achieved using a conventional timestamp alignment method in the field, eliminating the sampling delay differences between different sensors; Conventional filtering algorithms such as moving average filtering and median filtering are used to preprocess the synchronized multi-source data to remove impulse noise and high-frequency jump data, and output standardized preprocessed data. An adaptive fusion filtering algorithm adapted to the nonlinear motion model of the UAV (such as the Extended Kalman Filter (EKF) or Unscented Kalman Filter (UKF) algorithms commonly used in this field) is adopted to fuse the preprocessed multi-source data. The core fusion logic is as follows: the inertial data of the high-frequency inertial measurement unit is used to compensate for the low-frequency delay of the GPS positioning data, the barometric altimeter data is used to supplement the near-ground blind zone data for visual obstacle avoidance, and the cumulative drift of the inertial data is corrected by the GPS absolute position data. Finally, the core state parameters of the UAV with no noise and low latency are output.

[0032] A standardized finite state machine model is pre-built, based on relevant civil UAV control specifications and the physical flight characteristics of UAVs. The entire life cycle of the UAV is divided into standardized discrete states, including but not limited to: power-on standby, ground self-check, takeoff preparation, takeoff, hovering, cruise flight, maneuver control, landing, return to home, emergency response, and mission operation. Each discrete state has clearly defined entry conditions, exit conditions, and parameter safety thresholds. For example, the entry conditions for the takeoff state are a successful ground self-check, receipt of a valid takeoff command, and battery power meeting the minimum safety threshold. The exit condition is that the UAV reaches the set hovering altitude and switches to the hovering state. Based on the core state parameters output by the fusion filter, anomaly detection and state updates are performed in real time, outputting the precise flight status of the UAV.

[0033] A pre-built hierarchical safety rule expert system is constructed, including three categories: basic safety rules, scenario-based safety rules, and conflict resolution rules. Among them, basic safety rules are general rules that conform to the civil aviation UAV operation specifications and flight control basic safety logic, scenario-based safety rules are exclusive rules that can be customized according to different operation scenarios such as aerial photography, inspection, and plant protection, and conflict resolution rules are priority handling rules when multiple commands conflict. Meanwhile, a fixed command priority is defined in the expert system, with the priority from high to low as follows: emergency commands > safety commands > routine control commands > mission commands. Emergency commands include, but are not limited to, emergency stop, hovering, and return commands, to ensure that safety commands are executed first in emergency situations.

[0034] Using the verified control commands output from the source identification and verification step and the real-time flight status output from the real-time flight status perception step as inputs, conflict detection in three dimensions is completed in parallel: Timing conflict detection: Detects whether there is a timing logic conflict between the current instruction and a previous instruction that has not been completed. For example, if a takeoff instruction is received before the previous landing instruction has been completed, it is determined to be a timing conflict. State-command conflict detection: Detects whether the current command meets the entry / exit conditions and parameter safety thresholds corresponding to the current flight state. For example, if the drone is in ground standby mode and receives a descent command, it is determined to be a state-command conflict. Multi-command overlap conflict detection: Detects whether there is a risk of exceeding the flight safety boundary of the drone after the current command is overlapped with multiple commands executed in parallel. For example, if the drone is already at its maximum flight speed and receives a forward acceleration command, it is determined to be a multi-command overlap conflict.

[0035] The conflict detection across the three dimensions described above is executed in parallel. If any high-risk rule is triggered, the instruction is deemed illegal. Based on the detection results, a tiered execution decision is output. If a legitimate instruction does not trigger any violation rules, output a decision to allow execution. For instructions that trigger low-risk violations that do not directly cause safety incidents, output a decision to temporarily suspend execution, along with voice / visual prompts. For illegal instructions that trigger high-risk violations and will cause safety incidents upon execution, an interception + safety fallback action trigger decision is output. The safety fallback action includes at least one of hovering, returning to base, standing by, and deceleration. Those skilled in the art can configure the corresponding fallback action according to the conflict type.

[0036] Standardized upstream and downstream adaptation interfaces are pre-built. The upstream interface is an ASR engine adaptation interface, which is compatible with mainstream commercial online ASR engines and offline edge ASR engines. The downstream interface is a flight control adaptation interface, which is compatible with mainstream brand drone flight control SDKs and control terminals. Based on the execution decision output by the dynamic command legality reasoning steps, legal commands that are allowed are converted into a standard format that can be recognized by the corresponding flight controller through the downstream adaptation interface and sent to the flight controller for execution; high-risk and illegal commands are directly intercepted and not sent, and at the same time, the corresponding safety fallback actions are triggered. The system collects the execution results of flight control commands in real time, including whether the commands were successfully executed, changes in flight status after execution, and whether any abnormal situations occurred. The execution results are then fed back to the high-risk homophone feature library of the source identification and verification module and the safety rule expert system of the dynamic command legality reasoning module. This enables incremental updates to the feature library and rule library, continuously optimizing verification accuracy and adaptability, and forming a closed-loop control across the entire chain.

[0037] This embodiment uses an industry-level drone power line inspection scenario as an example to illustrate the complete implementation process of the present invention, as detailed below: Pre-configuration phase: For power line inspection scenarios, improve the high-risk homonym feature library and supplement high-risk commands specific to the inspection scenario, such as "approach the conductor / move away from the conductor" and "ascend / descend". At the same time, configure exclusive safety rules for power line inspection scenarios in the safety rule expert system, such as no-fly zone restrictions and safe distance restrictions from high-voltage conductors. Complete the upstream and downstream interface adaptation and connect to the corresponding ASR engine and the flight control SDK of the inspection drone.

[0038] During the flight operation phase: The operator issues control commands via voice. The general ASR engine outputs the Top-3 candidate recognition results. The system first determines whether the recognition result falls into the high-risk homonym feature library. If the recognition result contains homonyms such as "rising / falling", the lightweight model is activated for secondary verification and reordering, and the correct verified command is output.

[0039] The system simultaneously collects flight control data, obstacle avoidance radar data, and wind speed sensor data from the UAV, completes time synchronization and fusion filtering, and uses a finite state machine model to determine whether the UAV is currently in an inspection and cruise flight state.

[0040] Using the verified command and real-time flight status as inputs, the system performs conflict detection in three dimensions in parallel. If the verified command is "approach the wire", but the current distance between the drone and the wire is lower than the minimum safe distance set by the scenario-based safety rules, the system determines it as a state-command conflict, which is a high-risk illegal command, and outputs an interception decision.

[0041] The system directly intercepts the high-risk command without sending it to the flight controller, and simultaneously triggers a hovering safety fallback maneuver, issuing a voice prompt to the operator; at the same time, it feeds back the interception event and execution result to the rule base, completing incremental optimization of the rules and forming a complete closed-loop control.

Claims

1. A real-time security verification system for the entire voice control chain of a drone, comprising the steps of processing the voice recognition results output by a general-purpose ASR engine to obtain control commands and sending them to the drone flight control system for execution, characterized in that, It also includes the following steps: S1 Source Identification and Verification: A high-risk homophone feature library for UAVs with risk weight classification is pre-constructed. The feature library contains high-risk homophone pairs / word sets that are semantically contradictory in UAV operation scenarios and cause flight safety accidents after execution, as well as the acoustic features, text semantic features, and operation context features of the corresponding commands. Among them, high-risk contradictory commands are set as the highest verification weight. The system obtains the Top-N candidate recognition results, corresponding acoustic confidence scores, and contextual timing command information output by the general ASR engine. If the recognition result falls into the high-risk near-sound word feature library, a lightweight speech classification model adapted to the edge deployment is used to perform secondary classification and confidence reordering on the candidate recognition results, and outputs the verified control command with the highest confidence score that conforms to the control logic. If it does not fall into the library, the original recognition result is directly output as the verified control command. S2 Real-time Flight Status Perception: Real-time acquisition of UAV flight control core data and peripheral sensor data. The flight control core data includes altitude, three-axis velocity, three-axis attitude angle, GPS positioning data, battery power and voltage. The peripheral sensor data includes obstacle avoidance radar / visual obstacle avoidance distance data and wind speed sensor data. The multi-source data with different sampling frequencies is timestamped and synchronized. After filtering and preprocessing to remove impulse noise and high-frequency jump data, an adaptive fusion filtering algorithm adapted to the UAV's nonlinear motion model is used for fusion processing to output noise-free and low-latency core status parameters. Based on the civil aviation UAV control specifications and the physical characteristics of UAVs, a standardized finite state machine model is pre-constructed. The finite state machine model divides the flight state of the UAV throughout its entire life cycle into discrete states with clear entry / exit conditions and parameter safety thresholds. Based on the core state parameters, the model can determine and output the current precise flight state of the UAV in real time. S3 Dynamic Command Legality Reasoning: A pre-built hierarchical safety rule expert system is constructed. This system includes basic safety rules, scenario-based safety rules, and conflict resolution rules, and defines fixed command priorities, from highest to lowest: emergency commands > safety commands > routine control commands > mission commands. Using the verified control commands output in step S1 and the real-time flight status output in step S2 as inputs, conflict detection is performed in parallel across three dimensions: timing conflict detection, state-command conflict detection, and multi-command superposition conflict detection. Triggering any high-risk rule results in the command being deemed illegal. Based on the detection results, a tiered execution decision is output: legal commands are allowed execution; low-risk illegal commands are deferred with a warning; and high-risk illegal commands are intercepted with a safety fallback action trigger. S4 Full-link execution control and closed-loop feedback: Based on the execution decision output in step S3, legally granted instructions are converted into a format recognizable by the corresponding UAV flight control SDK and issued for execution; high-risk illegal instructions are directly intercepted and corresponding safety fallback actions are triggered; the execution results of flight control instructions are collected in real time and fed back to the high-risk homonym feature library in step S1 and the safety rule expert system in step S3, completing the incremental updates of the feature library and rule library, forming a full-link closed-loop control.

2. The method according to claim 1, characterized in that, In step S1, the construction of the high-risk homonym feature library also includes collecting environmental noise samples, multi-accent speech samples, and historical misidentification samples of the corresponding ASR engine in the target application scenario, extracting the corresponding features and supplementing them to the feature library.

3. The method according to claim 1, characterized in that, In step S2, the fusion processing logic of the adaptive fusion filtering algorithm is as follows: high-frequency inertial data is used to compensate for the low-frequency delay of GPS positioning data, altimeter data is used to supplement blind spot data for visual obstacle avoidance, and absolute position data is used to correct the cumulative drift of inertial data.

4. The method according to claim 1, characterized in that, In step S3, the safety fallback action includes at least one of hovering, returning to base, staying in place, and deceleration.

5. The method according to claim 1, characterized in that, In step S1, the general ASR engine includes a commercial online ASR engine and an offline edge ASR engine.

6. The method according to claim 1, characterized in that, In step S2, the peripheral sensor data also includes ambient light data, temperature and humidity data, and air pressure data.

7. The method according to claim 1, characterized in that, In step S3: The timing conflict detection is to detect whether there is a timing logic conflict between the current instruction and the preceding instruction that has not been completed. The state-command conflict detection is used to detect whether the current command meets the entry / exit conditions and parameter safety thresholds corresponding to the current flight state. The multi-instruction superposition conflict detection is used to detect whether there is a risk of exceeding the flight safety boundary of the UAV after the current instruction is superimposed with multiple instructions executed in parallel.

8. The method according to claim 1, characterized in that, In step S4, at least two mainstream commercial ASR engines are connected through standardized upstream interfaces, and at least two mainstream brand drone flight control SDKs are connected through standardized downstream interfaces.

9. A real-time security verification system for the entire voice control chain of unmanned aerial vehicles (UAVs), characterized in that, include: The source identification and verification module is used to pre-build a feature library of high-risk homophones for drones with risk weight classification, obtain the identification-related data output by the general ASR engine, complete the secondary verification and error correction of high-risk homophones, and output the verification control command. The flight status perception module is used to collect multi-source flight data of UAVs, complete data synchronization, preprocessing and fusion filtering, and judge and output the current flight status of UAVs in real time based on a pre-built standardized finite state machine model. The dynamic command legality reasoning module is used to pre-build a hierarchical safety rule expert system. It takes the verified control commands and real-time flight status as inputs, completes parallel conflict detection in three dimensions, and outputs hierarchical command execution decisions. The end-to-end execution control and closed-loop feedback module is used to issue or intercept commands based on execution decisions, while collecting flight control execution results and feeding them back to the corresponding modules to complete incremental updates of the feature library and rule library, thereby achieving end-to-end closed-loop control.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-8.

Images