UAV flight safety control method based on visual disturbance and trajectory deviation coupling

Abstract

This invention provides a UAV flight safety control method based on the coupling of visual perturbation and trajectory deviation, belonging to the field of unmanned system testing and verification technology. The method includes: S1, setting a set of UAV flight expectation compliance assessment thresholds; S2, constructing a UAV visual adversarial test scenario library; S3, injecting data and model dual-sensitivity perturbations into the UAV visual perception and control model; S4, determining the reliability assessment parameters of UAV dynamic evaluation and physical execution; S5, controlling the UAV physical actuators according to the reliability assessment parameters of the UAV flight control system. This invention effectively determines the robustness boundary of the model under input perturbations through the data and model dual-sensitivity perturbation injection mechanism; uses the expectation compliance parameters fused by multi-dimensional features to provide timely danger warnings; and establishes a hierarchical adaptive physical takeover mechanism to effectively prevent loss of control and crash accidents caused by algorithm logic collapse.

Description

Technical Field

[0001] This invention relates to the field of unmanned system testing and verification technology, specifically to a method for unmanned aerial vehicle (UAV) flight safety control based on the coupling of visual disturbance and trajectory deviation. Background Technology

[0002] With the development of artificial intelligence, drones are widely used in complex scenarios such as urban logistics and autonomous inspection. However, compared with traditional avionics systems, systems equipped with deep learning algorithms exhibit drastically different failure characteristics, posing serious challenges to existing reliability assessment methods: First, high input sensitivity: the algorithm is extremely sensitive to input data. Traditional methods focus on hardware failures, but the system may make identification errors due to slight changes in lighting, texture noise, or adversarial samples. These pixel-level perturbations are often ignored in traditional testing. Second, output unpredictability and model black-box uncertainty: deep neural networks lack physical interpretability; it is difficult to characterize the activation state of neurons within the model, and it is difficult to predict the system's output behavior in corner cases. Third, risk masking effect: traditional assessments are mostly based on the statistical failure rate (MTBF). Before the system completely fails, there are often high-confidence misjudgments or hesitations, and high-entropy evaluation anomalies. Existing methods are unable to capture such early evaluation layer deviations, leading to sudden catastrophic accidents when the system is considered to be in a normal state.

[0003] Therefore, this invention aims to propose a verification framework that no longer relies on traditional hardware failure rates but is based on expected compliance, in order to solve the problem of difficulty in measuring the reliability of aerospace equipment systems in complex scenarios. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention aims to provide a UAV flight safety control method based on the coupling of visual perturbation and trajectory deviation. Through a dual-sensitivity perturbation injection mechanism involving both data and the model, adversarial noise is injected along the gradient direction in the data space, and a random mask is injected in the parameter space, effectively determining the robustness boundary of the model under input perturbations. An expected compliance parameter based on multi-dimensional feature fusion is constructed, introducing algorithm confidence and information entropy. An evaluation function integrating physics and evaluation is built and nonlinearly aggregated with physical execution error, immediately issuing an alarm when blind confidence or high-entropy hesitation occurs within the algorithm. A hierarchical adaptive physical takeover mechanism based on compliance metrics is established, improving the actual flight safety of the UAV. Closed-loop control from the algorithm layer to the physical execution layer (motors / servos) effectively prevents loss of control and crashes caused by algorithm logic collapse.

[0005] Specifically, the present invention provides a method for unmanned aerial vehicle (UAV) flight safety control based on the coupling of visual disturbance and trajectory deviation, which includes the following steps: S1: Obtain the drone's flight speed in the current flight mission. Euclidean distance deviation of trajectory and motor speed control signal Generate a set of thresholds for assessing the compliance of drone flight expectations. ; S2: Acquire raw images taken by the drone in clear weather. ; Construct a joint loss function to train the generative adversarial network; The minimax game objective function of the generative adversarial network; Define the total loss function of the generator. Optimize the output adversarial test dataset ; S3: Obtain the adversarial test scenario library generated in step S2, and input the adversarial test dataset. Images in and parameters of deep neural network visual perception model The objective function for constructing a visual perception model Calculate the gradient of the model loss function with respect to the input image. Generate perturbation samples with injected adversarial features. Output system evaluation execution state variables and expected physical location ; S4: Obtain the system evaluation execution state variables from S3 Determine the Euclidean distance deviation of the UAV trajectory Obtain the set of drone flight expectation compliance assessment thresholds in S1. Maximum Predictive Information Entropy By combining the confidence level and information entropy of the perception model with the underlying physical trajectory deviation, a nonlinear aggregation formula is used to determine the reliability assessment parameters of the UAV flight control system. ; S5: Reliability evaluation parameters of the UAV flight control system obtained from S4 Perform three-level UAV flight status assessment; compare and assess the reliability evaluation parameters of the UAV flight control system. The system determines the area in which the drone is located and controls the drone's physical actuators to perform actions.

[0006] Preferably, the set of drone flight expectation compliance assessment thresholds in step S1 for: ; in, This is a set of thresholds for assessing the compliance of drone flight expectations. For safety confidence threshold; The maximum predictive information entropy; This represents the maximum Euclidean deviation allowed by the physical trajectory. This represents the maximum safe flight speed allowed for the drone in the current obstacle avoidance scenario. To ensure the stability of the drone's attitude, the motor speed control signal is set within a safe range.

[0007] Preferably, the minimax game objective function for generating the adversarial network in step S2 is as follows: ; in, To determine the sign of the minimum value; To determine the sign of the maximum value; Output the result of the game objective function; For discriminators; For generator; The expected function of the original image; It is a logarithmic function; The discriminator outputs the result for the original image; The generator outputs the result of the original image; This is the original image.

[0008] Preferably, the total loss function of the generator in step S2 Specifically: ; in: The total loss function of the generator; To generate an adversarial loss function; For image physical consistency constraints; This is the confidence output function for the visual recognition algorithm of the drone; To counter the loss items of attacks; The first hyperparameter is used to balance all terms; This is the second hyperparameter for balancing the weights.

[0009] Preferably, the objective function of the visual perception model in step S3 is... Specifically: ; in, The objective function of the visual perception model; This represents the total number of obstacle categories. Index for obstacle categories; Encoding of the actual label; To predict the category to which the target belongs in a perception network The probability of; These are the parameters for a deep neural network visual perception model. For adversarial test datasets The image in; For adversarial test datasets The distance to obstacles in the middle.

[0010] Preferably, the perturbation sample for injecting adversarial features in step S3 is: ; in, Perturbation samples for injecting adversarial features; For resistance strength coefficient; It is a symbolic function; The gradient of the input image.

[0011] Preferably, the system evaluation execution state variables in step S3 Specifically, it includes: ; ; in, The maximum classification confidence; To predict information entropy; This is the predicted probability distribution vector; The predicted probability distribution vector The first in Each element value; Index of the predicted probability distribution vector elements; This is a time parameter.

[0012] Preferably, the Euclidean distance deviation of the UAV trajectory in step S4 is: ; in, The Euclidean distance deviation of the drone trajectory; The expected physical location of the drone; This refers to the actual physical location of the drone.

[0013] Preferably, the reliability evaluation parameters of the UAV flight control system in step S4 for: ; in, These are parameters for reliability evaluation of UAV flight control systems. The weight for the highest classification confidence; The maximum classification confidence; For natural index parameters; The maximum predictive information entropy; It is the tangent function; This is the Euclidean distance deviation coefficient for the trajectory.

[0014] Preferably, in step S5, the reliability evaluation parameters of the UAV flight control system are used. The third-level drone flight status assessment is performed as follows: Level 1 flight status: The UAV flight control system is functioning normally, outputting standard pulse width modulation (PWM) signals to drive the motors and execute automatic control path planning. Level 2 flight status: In the ambiguous assessment zone, reduce flight speed. The flight control system reduces the average PWM duty cycle sent to the four motors while increasing the fusion weight of the LiDAR data. Level 3 flight status: If the expected deviation zone is detected, automatic control is deemed to have failed; visual navigation commands are immediately cut off, and the flight control system switches to inertial hold mode.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention proposes a dual-sensitivity perturbation injection mechanism for data and model, which effectively determines the robustness boundary of the algorithm. Unlike traditional electromagnetic or physical fault injection, this invention targets the characteristics of artificial algorithms, injects adversarial noise along the gradient direction in the data space, and injects random masks in the parameter space, effectively determining the robustness boundary of the model under input perturbation; filling the gap in traditional testing methods that are difficult to assess the black-box risk of algorithms.

[0016] (2) This invention constructs an expected compliance parameter based on multi-dimensional feature fusion, abandons the single physical error parameter, introduces algorithm confidence and information entropy, constructs an evaluation function that integrates physics and evaluation, and performs nonlinear aggregation with physical execution error; this expected compliance parameter can immediately issue an alarm when the physical position of the UAV has not deviated from the safe zone, but the algorithm has already shown blind confidence or high entropy hesitation, thus solving the timeliness problem of traditional methods that only report errors after a collision.

[0017] (3) The present invention establishes a hierarchical adaptive physical takeover mechanism based on compliance measurement, which improves the actual flight safety of UAVs. Existing technologies often lack a smooth transition strategy when an anomaly is detected, while the present invention divides the flight state into three levels: normal, evaluation ambiguity, and expected deviation based on the expected compliance index. In the evaluation ambiguity stage, the motor speed signal is reduced to actively decelerate in order to gain calculation time. In the expected deviation stage, the visual navigation command is immediately cut off, the servo is locked, and the vertical hovering or landing mode based on inertial navigation is switched. This closed-loop control from the algorithm layer to the physical execution layer motor / servo effectively prevents uncontrolled crashes caused by algorithm logic collapse. Attached Figure Description

[0018] Figure 1 This is a control block diagram of the UAV flight safety control method based on the coupling of visual disturbance and trajectory deviation according to the present invention. Figure 2 This is the overall logical flowchart of the REACH framework of this invention; Figure 3 A time-series response graph showing the relationship between scene adversarial intensity, model uncertainty entropy, and physical trajectory deviation; Figure 4 This is a comparison curve of the conformity index of this invention and the traditional MTBF evaluation results; Figure 5 This is a schematic diagram illustrating the evaluation of the failure threshold in an embodiment of the present invention; Figure 6 This is a schematic diagram of the physical hysteresis effect in an embodiment of the present invention. Detailed Implementation

[0019] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0020] This invention proposes a method for unmanned aerial vehicle (UAV) flight safety control based on the coupling of visual disturbance and trajectory deviation, such as... Figure 1 As shown, a set of thresholds for evaluating the compliance of UAV flight expectations is set; a library of UAV visual adversarial test scenarios is constructed; data and model dual-sensitivity perturbations are injected into the UAV visual perception and control model; reliability evaluation parameters for UAV dynamic evaluation and physical execution are determined; and the UAV physical actuators are controlled according to the reliability evaluation parameters of the UAV flight control system. This embodiment of the invention uses flight speed... (m / s), trajectory Euclidean distance deviation (m), and motor speed control signal Three variables serve as the benchmark for UAV flight status and execution evaluation, and are deeply integrated into the calculation process of scene generation and model closed-loop recovery. Through the REACH framework, a verification chain is constructed from the setting of the flight control system's operational design domain (ODD) to closed-loop recovery, enabling precise measurement of flight control system evaluation and execution; such as... Figure 2 As shown, the specific steps include: S1: Set a set of thresholds for drone flight expectation compliance assessment Obtain the drone's flight speed during the current flight mission. Euclidean distance deviation of trajectory and motor speed control signal The safe operating benchmarks are established, and based on the UAV flight mission profile, the operational design domain (ODD) of the flight control system is set to determine the expected compliance boundaries of key UAV flight performance parameters. These expected compliance boundaries include the confidence threshold of the visual algorithm output, the upper limit of the decision entropy value, the physical trajectory tracking error tolerance, and the control constraints of the underlying motors. In this embodiment, the UAV is configured to navigate between urban buildings, relying on visual sensors for simultaneous localization and mapping (SLAM) navigation and obstacle avoidance. In this embodiment of the invention, the operating design domain (ODD) is set as follows: light intensity 10-10000 lux, relative wind speed <10 m / s.

[0021] The expected compliance parameters are as follows: within the operating design domain (ODD), the confidence level of the UAV visual recognition algorithm should be... The Euclidean distance deviation of trajectory tracking should satisfy... And flight speed With motor control signals It must be kept within the set safety threshold range.

[0022] The input variables are: a preset flight mission, including a city building navigation scenario and the physical and dynamic constraints of the drone. The execution process involves setting the physical boundaries for safe drone flight, including: speed boundaries, trajectory deviation boundaries, and motor speed boundaries; the evaluation boundaries are visual confidence and decision entropy.

[0023] Output variables: Set of thresholds for assessing the compliance of drone flight expectations for:

[0024] ; in, This is a set of thresholds for assessing the compliance of drone flight expectations. For safety confidence threshold; The maximum predictive information entropy represents the upper limit of uncertainty; The maximum allowable Euclidean deviation of the physical trajectory, i.e., the physical variable. The safety limit; This represents the maximum safe flight speed allowed for the drone in the current obstacle avoidance scenario, i.e., the safe upper limit of the physical variable v. To ensure the safe range of motor speed control signals for maintaining the drone's attitude stability, i.e., physical variables The scope of constraints.

[0025] S2: Construct a test scenario library for UAV visual adversarial capabilities. Receive raw images collected by the UAV's visual sensors under normal conditions, and based on a data-driven method, extract high-dimensional data features that are prone to causing visual algorithm failures in the flight control system, constructing a test scenario library containing boundary scenes and out-of-distribution data. In this embodiment, a generative adversarial network (GAN) is used to generate long-tail scene data, such as high-dimensional feature samples like strong reflections from glass curtain walls, low contrast caused by haze, and dynamic obstacle occlusion, which are then stored in a knowledge base. The input variable for this step is the raw image collected by the UAV under clear weather conditions. .

[0026] Using a Generative Adversarial Network (GAN) generator on the original image Perform pixel-level high-dimensional feature perturbation to generate corresponding adversarial examples. To ensure that the generated edge scenes conform to real physical laws while effectively triggering the failure of UAV visual evaluation, this invention constructs a joint loss function to train the Generative Adversarial Network (GAN); the specific process is as follows: Use generator G to receive the original image and output adversarial examples. To simulate pixel nonlinear shifts caused by strong backlighting and dynamic blurring caused by high-speed flight, a discriminator D is set to distinguish between real, harsh environment images and generated images; the minimax game objective function of the generative adversarial network (GAN) is: ; in, To determine the sign of the minimum value; To determine the sign of the maximum value; Output the result of the game objective function; For discriminators; For generator; The expected function of the original image; It is a logarithmic function; The discriminator outputs the result for the original image; The generator outputs the result of the original image; This is the original image.

[0027] To achieve efficient perturbation generation in UAV flight control vision algorithms, the total loss function of generator G is defined. for: ; in: Let G be the total loss function of the generator; To generate an adversarial loss function and ensure the realism of generated scenes such as lighting and fog; This is an image physical consistency constraint term used to limit the perturbation amplitude and prevent excessive image distortion from exceeding physical norms. This is the confidence output function for the visual recognition algorithm of the drone; The security confidence threshold set in step S1; To counteract the attack loss term, it is used to force the output confidence of the UAV vision algorithm to fall below a safety threshold, thereby inducing evaluation failure; The first hyperparameter is used to balance all terms; This is the second hyperparameter for balancing the weights.

[0028] The output variable is the adversarial test dataset. ,in, An adversarial image infused with the perturbation calculated using the above formula. , This step outputs the adversarial test dataset, which contains the ground truth labels for the corresponding obstacle distances. A library of adversarial test scenarios is created and used directly as the input source for the S3 model.

[0029] S3: Inject data and model dual-sensitivity perturbations into the UAV visual perception and control model.

[0030] The adversarial test scenario library generated in step S2 is obtained. The impact of environmental feature degradation on the visual evaluation module of the UAV flight control system is simulated, and a dual perturbation injection is performed, incorporating the instability within the human-machine visual perception and control model. The injected adversarial perturbation framework is as follows: noise from the algorithm's feature space is superimposed on the image input. Simultaneously, random dropout is applied to the fully connected layers of the neural network to simulate the internal parameter instability of the UAV visual perception and control model when processing out-of-distribution data. The input variable is the adversarial test dataset. Images in And the parameters of the deep neural network visual perception model pre-deployed within the UAV flight control system The specific execution process is as follows: Adding input perturbations: Constructing the objective function of the visual perception model before calculating the gradient. Specifically: ; in, The objective function of the visual perception model; This represents the total number of obstacle categories. Index for obstacle categories; Encoding of the actual label; To predict the category to which the target belongs in a perception network The probability of; It is a logarithmic function; These are the parameters for a deep neural network visual perception model. For adversarial test datasets The image in; For adversarial test datasets The distance to obstacles in the middle.

[0031] Based on the objective function of the visual perception model, calculate the gradient of the model's loss function with respect to the input image. And generate perturbation samples with injected adversarial features as follows: ; in, Perturbation samples for injecting adversarial features; For resistance strength coefficient; It is a symbolic function; The gradient of the input image.

[0032] During inference, a Dropout mask M is applied to the fully connected layers of the UAV visual perception model, such as a deep convolutional obstacle recognition network, to randomly deactivate some neurons, thus perturbing the model. The perturbed samples with injected adversarial features are then used to perform this perturbation. The input is fed into the masked UAV visual perception model for feature extraction and classification calculation to obtain the predicted probability distribution vector. The forward propagation process is then executed. Subsequently, the features of the perception results are passed to the downstream UAV trajectory planning model, such as the local path planner. The output variables are: system evaluation execution state variables. Includes: maximum classification confidence Predicting information entropy , used for S4 calculation, where, The above predicted probability distribution vector The value of the i-th element in the model; the expected physical location planned by the downstream UAV trajectory planning model at the next moment based on the current disturbance assessment state. Used in S4 to calculate the Euclidean distance deviation of the trajectory by combining real physical coordinates. .

[0033] S4: Determine the reliability assessment parameters for UAV dynamic evaluation and physical execution. .

[0034] Combined with the set of drone flight expectation compliance assessment thresholds set in step S1 The system monitors the perception confidence, output entropy, and physical execution deviation of the flight control system in real time under anti-disturbance conditions. A nonlinear aggregation algorithm is used to calculate the expected compliance parameters of the flight control system, i.e., reliability assessment parameters. This is to quantify the degree to which the actual flight behavior of drones matches the design expectations.

[0035] The input variables are the system evaluation execution state variables in S3. Including maximum classification confidence Predictive information entropy The expected physical location of the model planning The actual physical location of the drone comes from its GPS or RTK (Real-Time Dynamic Differential) module. ; Set of drone flight expectation compliance assessment thresholds from S1 Maximum Predictive Information Entropy .

[0036] Based on the expected physical location and the actual location of the drone, the Euclidean distance deviation of the drone's trajectory at the current moment is calculated as follows: ; in, The Euclidean distance deviation of the drone trajectory; The expected physical location planned for the model; The actual physical location of the drone; like Figure 3 The figure shows the time-series response relationship between scene adversarial strength, model uncertainty entropy, and physical trajectory deviation. Combining the confidence level of the perception model, information entropy, and the underlying physical trajectory deviation, the following nonlinear aggregation formula is used to calculate the reliability assessment parameters of the UAV flight control system at the current moment. for: ; in, These are parameters for reliability evaluation of UAV flight control systems. The weight for the highest classification confidence; The maximum classification confidence; For natural index parameters; To predict information entropy; It is the tangent function; This is the Euclidean distance deviation coefficient for the trajectory.

[0037] S5: Based on the reliability evaluation parameters of the UAV flight control system Control the physical actuators of the drone.

[0038] The reliability evaluation parameters of the UAV flight control system calculated based on step S4 The system continuously assesses the status of the UAV flight control system and, when it determines that there are uncontrollable risks in visual assessment or trajectory execution, directly intervenes in the underlying flight speed and motor speed control signals of the UAV, triggering a degradation switch to deterministic safety rule control. The input variables for this step are: real-time reliability assessment parameters of the UAV flight control system. Compare and evaluate the reliability assessment parameters of the UAV flight control system. The area in which it is located, and controls the physical actuators of the drone.

[0039] Level 1 flight status: The system is functioning normally. The flight controller outputs a standard pulse width modulation (PWM) signal to drive the motor and execute the automatic control planning path.

[0040] Level 2 flight status: Assessment of ambiguous areas: Reduce flight speed to The specific operation involves: the flight control system reducing the average PWM duty cycle sent to the four motors, while simultaneously increasing the fusion weight of the LiDAR data.

[0041] Level 3 flight status: Expected deviation zone: Automatic control failure is determined. Visual navigation commands are immediately discontinued, and the flight control system switches to inertial hold mode.

[0042] Physical action: Lock the drone's attitude angle Roll / Pitch to 0 via the controller and adjust the motor speed to achieve vertical acceleration. Perform stationary hovering or The aircraft descends vertically at high speed to prevent it from crashing.

[0043] Action triggering in this embodiment of the invention: Based on the reliability assessment parameters of the UAV flight control system =0.18 < 0.3, the flight control system determines that the current environment has exceeded the effective domain of the algorithm. For example... Figure 4 The figure shows a comparison curve between the conformity index of this invention and the traditional MTBF evaluation results; the execution recovery of this invention embodiment: immediately disable visual navigation commands, switch to instrument landing logic, and rely solely on the anti-interference laser rangefinder and inertial measurement unit (IMU) to maintain attitude, vertically climb or hover, avoiding crashes caused by discrepancies between the algorithm and expectations. As can be seen from the above embodiments, this invention can effectively verify the true reliability of the flight control system under black-box and highly sensitive characteristics. This formula vividly demonstrates how this invention, by capturing the psychological fluctuations, confidence levels, and entropy of the AI ​​algorithm before a physical accident occurs, provides early warning of flight control system failure, filling the gap in existing testing methods in the artificial domain. Figure 5 The diagram illustrates the assessment of the critical failure point, showing that although the physical location has not yet deviated, the entropy value spikes prematurely, and the warning level drops, demonstrating the UAV flight control system's advanced perception of potential accidents. Figure 6 The diagram shows the physical hysteresis effect, illustrating that traditional UAV flight control systems only issue an alarm when an actual collision occurs and the distance is zero, while the present invention triggers takeover in advance.

[0044] The beneficial effects of this invention are as follows: This invention proposes a UAV flight safety control method based on the coupling of visual disturbances and trajectory deviations. It introduces a dual-sensitivity disturbance injection mechanism for both data and models, effectively determining the robustness boundary of the algorithm. Unlike traditional electromagnetic or physical fault injection, this invention, targeting the characteristics of artificial algorithms, injects adversarial noise along the gradient direction in the data space and a random mask in the parameter space, effectively determining the robustness boundary of the model under input perturbations. It fills the gap in traditional testing methods for assessing the black-box risk of algorithms. Based on the expected compliance parameters fused from multi-dimensional features, it constructs an evaluation function that integrates physics and evaluation, enabling early warning before the UAV's physical position deviates from the safe zone, solving the timeliness problem of traditional methods that only report errors after a collision. It establishes a hierarchical adaptive physical takeover mechanism based on compliance metrics, classifying flight states into three levels: normal, ambiguous evaluation, and expected deviation, according to the expected compliance index. Closed-loop control from the algorithm layer to the physical execution layer (motors / servos) effectively prevents loss of control and crashes caused by algorithm logic collapse.

[0045] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for unmanned aerial vehicle (UAV) flight safety control based on the coupling of visual disturbance and trajectory deviation, characterized in that: It includes: S1: Obtain the drone's flight speed in the current flight mission. Euclidean distance deviation of trajectory and motor speed control signal Generate a set of thresholds for assessing the compliance of drone flight expectations. ; S2: Acquire raw images taken by the drone in clear weather. ; Construct a joint loss function to train the generative adversarial network; The minimax game objective function of the generative adversarial network; Define the total loss function of the generator. Optimize the output adversarial test dataset ; S3: Obtain the adversarial test scenario library generated in step S2, and input the adversarial test dataset. Images in Parameters of deep neural network visual perception models The objective function for constructing a visual perception model Calculate the gradient of the model loss function with respect to the input image. Generate perturbation samples with injected adversarial features. ; Output system evaluation execution state variables and expected physical location ; S4: Obtain the system evaluation execution state variables from S3 Determine the Euclidean distance deviation of the UAV trajectory Obtain the set of drone flight expectation compliance assessment thresholds in S1. Maximum Predictive Information Entropy By combining the confidence level and information entropy of the perception model with the underlying physical trajectory deviation, a nonlinear aggregation formula is used to determine the reliability assessment parameters of the UAV flight control system. ; S5: Reliability evaluation parameters of the UAV flight control system obtained from S4 Perform three-level UAV flight status assessment; compare and assess the reliability evaluation parameters of the UAV flight control system. The system determines the area in which the drone is located and controls the drone's physical actuators to perform actions.

2. The UAV flight safety control method based on the coupling of visual disturbance and trajectory deviation according to claim 1, characterized in that: The set of drone flight expectation compliance assessment thresholds in step S1 for: ； in, This is a set of thresholds for assessing the compliance of drone flight expectations. The safety confidence threshold; The maximum predictive information entropy; This represents the maximum Euclidean deviation allowed by the physical trajectory. This represents the maximum safe flight speed allowed for the drone in the current obstacle avoidance scenario. To ensure the stability of the drone's attitude, the motor speed control signal is set within a safe range.

3. The UAV flight safety control method based on the coupling of visual disturbance and trajectory deviation according to claim 1, characterized in that: Step S2 generates the minimax game objective function for the adversarial network, specifically as follows: ； in, To determine the sign of the minimum value; To determine the sign of the maximum value; Output the result of the game objective function; For discriminators; For generator; The expected function of the original image; It is a logarithmic function; The discriminator outputs the result for the original image; The generator outputs the result of the original image; This is the original image.

4. The UAV flight safety control method based on the coupling of visual disturbance and trajectory deviation according to claim 1, characterized in that: The total loss function of the generator in step S2 Specifically: ； in: The total loss function of the generator; To generate an adversarial loss function; For image physical consistency constraints; This is the confidence output function for the visual recognition algorithm of the drone; To counter the loss items of attacks; The first hyperparameter is used to balance all terms; This is the second hyperparameter for balancing the weights.

5. The UAV flight safety control method based on the coupling of visual disturbance and trajectory deviation according to claim 1, characterized in that: The objective function of the visual perception model in step S3 Specifically: ； in, The objective function of the visual perception model; This represents the total number of obstacle categories. Index for obstacle categories; Encoding of the actual label; To predict the category to which the target belongs in a perceptual network The probability of; These are the parameters of a deep neural network visual perception model; For adversarial test datasets The image in; For adversarial test datasets The distance to obstacles in the middle.

6. The UAV flight safety control method based on the coupling of visual disturbance and trajectory deviation according to claim 1, characterized in that: The perturbation samples for injecting adversarial features in step S3 are: ； in, Perturbation samples for injecting adversarial features; For resistance strength coefficient; It is a symbolic function; The gradient of the input image.

7. The UAV flight safety control method based on the coupling of visual disturbance and trajectory deviation according to claim 1, characterized in that: The system evaluation execution state variables in step S3 Specifically, it includes: ； ； in, The maximum classification confidence; To predict information entropy; This is the predicted probability distribution vector; The predicted probability distribution vector The first in Each element value; Index of the predicted probability distribution vector elements; This is a time parameter.

8. The UAV flight safety control method based on the coupling of visual disturbance and trajectory deviation according to claim 1, characterized in that: The Euclidean distance deviation of the UAV trajectory in step S4 is: ； in, The Euclidean distance deviation of the drone trajectory; The expected physical location of the drone; This refers to the actual physical location of the drone.

9. The UAV flight safety control method based on the coupling of visual disturbance and trajectory deviation according to claim 1, characterized in that: Reliability evaluation parameters of the UAV flight control system in step S4 for: ； in, These are parameters for reliability evaluation of UAV flight control systems. The weight for the highest classification confidence; The maximum classification confidence; For natural index parameters; The maximum predictive information entropy; It is the tangent function; This is the Euclidean distance deviation coefficient for the trajectory.

10. The UAV flight safety control method based on the coupling of visual disturbance and trajectory deviation according to claim 1, characterized in that: In step S5, the reliability evaluation parameters of the UAV flight control system are used. The third-level drone flight status assessment is performed as follows: Level 1 flight status: The UAV flight control system is functioning normally, outputting standard pulse width modulation (PWM) signals to drive the motors and execute automatic control path planning. Level 2 flight status: In the ambiguous assessment zone, reduce flight speed. ; The flight control system reduces the average PWM duty cycle sent to the four motors, while increasing the fusion weight of the LiDAR data. Level 3 flight status: ; If the expected deviation occurs, the automatic control system is deemed to have failed; Immediately disengage visual navigation commands and switch the flight control system to inertial hold mode.

Images