Aircraft active safety rolling guidance method based on Gaussian process

Abstract

The application discloses a kind of active safety rolling guidance method of aircraft based on Gaussian process, belong to aircraft navigation, guidance and control field.The application uses Gaussian process regression algorithm with space-time information entropy adaptive sampling, can optimize hyperparameter on line under limited computing power, accurately output the future trajectory mean and position variance of obstacle, improve the precision and efficiency of active safety rolling guidance of aircraft.The risk field index of the application is coupled with spatial distance, relative velocity and prediction variance, when facing high-speed approach or high-uncertainty obstacle, the application can automatically inflate safety anti-collision ellipsoid, provide sufficient safety margin.The application introduces risk index attenuation mechanism in NMPC cost function.In high danger, automatically reduce the tracking weight of global path profile, preferentially guarantee obstacle avoidance, while driving head actively aiming at obstacle, ensure that perception line of sight is not lost, form the bionic intelligent behavior of "more dangerous, more attention", realize safety rolling guidance.

Description

Technical Field

[0001] This invention belongs to the field of aircraft navigation, guidance and control technology, and specifically relates to an active safety roll guidance method for aircraft based on Gaussian process regression and nonlinear model predictive control in dynamic uncertain environments. Background Technology

[0002] In applications such as inspection, search and rescue, urban low-altitude navigation, and emergency logistics, aircraft often need to complete autonomous flight missions from a starting point to a target point in complex environments with dynamic obstacles such as pedestrians, vehicles, and other non-cooperative aircraft or mobile devices. Compared with static obstacles, dynamic obstacles are characterized by changes in motion intent over time, high maneuverability, and greater uncertainty due to observation noise and obstruction. This requires the aircraft's trajectory planning not only to meet the efficiency of reaching the target, but also to maintain a sufficient safe distance and have high-frequency real-time adjustment capabilities throughout the flight.

[0003] In existing dynamic obstacle avoidance and local path planning methods, one type of method typically performs regular obstacle avoidance based on the instantaneous observed position of the obstacle, or uses simple uniform velocity or uniform acceleration extrapolation to predict the short-term motion trend of the obstacle. This type of method is simple to implement, but when the obstacle's motion pattern changes, there is noise in the observation, or there is a short-term visual loss, the prediction error of the kinematic extrapolation can easily accumulate rapidly, leading to serious problems such as insufficient safety margin in the planning results, violent path oscillations, or even collisions.

[0004] To overcome the shortcomings of simple extrapolation, another approach employs nonlinear model predictive control (MMDC) to perform rolling optimization of the control input sequence within a given prediction time domain. This method comprehensively considers path tracking, arrival efficiency, and obstacle avoidance safety under physical constraints, exhibiting strong online planning capabilities. However, traditional MMDC in dynamic environments typically relies heavily on deterministic obstacle prediction trajectories or uses only a fixed safety expansion radius. This approach struggles to explicitly and quantitatively incorporate the uncertainty of dynamic obstacle predictions into the safety decision layer, leading to a dilemma of insufficient or excessive conservatism when uncertainty is high.

[0005] Furthermore, during actual flight, aircraft are subject to strict underlying physical constraints such as thrust limitations and attitude angle deflection rates. If the cost function or state constraint design at the planning layer is unreasonable, the solution of nonlinear model predictive control is prone to abrupt changes in control input, causing the aircraft's motion to be unsmooth or even directly exceeding the physical limits of the actuators. At the same time, extreme control commands can also easily lead to the failure or timeout of the online optimizer, reducing the reliability of the system.

[0006] More importantly, existing local planning methods often treat geometric obstacle avoidance and sensor field-of-view coverage separately. Accurate prediction of dynamic obstacles heavily relies on continuous and effective observation from onboard sensing sensors. When obstacles are at the edge of the field of view or in obstructed areas, prediction uncertainty increases significantly. Traditional planning strategies, when performing aggressive evasive maneuvers, often only change the aircraft's three-dimensional position, which can easily cause dangerous obstacles to slide out of the onboard sensor's field of view. If the planning engine does not incorporate sensor field of view and aircraft heading adjustment into a unified objective, the system can only passively respond to unknown risks and will find it difficult to improve the observation quality of high-risk obstacles by actively adjusting yaw attitude or line-of-sight.

[0007] Therefore, current technology urgently needs an online trajectory planning method for active safety of aircraft in dynamic and uncertain environments. This method should be able to deeply integrate the accurate prediction of the future nonlinear motion of dynamic obstacles and the quantitative assessment of their uncertainties within a rolling optimization framework, thereby constructing a dynamic and adaptive safety collision avoidance mechanism. It should also take into account the smooth executability of the underlying control and the need for active perception, so that the aircraft can not only effectively and safely avoid dangers, but also continuously observe and track dangerous obstacles, thus comprehensively improving the safety, robustness and feasibility of aircraft in dynamic and complex environments. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide an active safety roll guidance method for aircraft based on Gaussian processes. This method uses Gaussian processes to infer the uncertainty of obstacles online, constructs a unified dynamic risk field, and achieves deep coupling and coordinated control of safe collision avoidance, accurate tracking, and active perception within the framework of model predictive control (NMPC).

[0009] The objective of this invention is achieved through the following technical solution:

[0010] This invention discloses an active safety roll guidance method for aircraft based on Gaussian processes, comprising the following steps:

[0011] Step 1: Establish the dynamic model of the aircraft.

[0012] Step 2: During the flight of the aircraft, establish a training set sequence of historical effective trajectory points based on adaptive sampling of spatiotemporal information entropy; use the training set to optimize the set of hyperparameters of the kernel function through online maximum likelihood estimation; calculate the mean vector of the predicted position and its position uncertainty variance in the future prediction time domain of the dynamic obstacle using the Gaussian process posterior inference formula.

[0013] Step 3: Based on the dynamic model in Step 1 and the predicted position mean vector and position uncertainty variance output in Step 2, establish a unified dynamic risk field that combines spatial Euclidean distance, relative velocity, and comprehensive uncertainty variance.

[0014] Step 4: Based on the unified dynamic risk field described in Step 3, establish a safety boundary constraint obstacle function with a dynamic adaptive expansion mechanism, a trajectory contour error cost function based on risk adaptive adjustment, an active perception value field cost function considering line-of-sight occlusion, and a heading alignment target cost function; introduce a target point attraction cost function and a control input smoothing cost function to construct a comprehensive cost function.

[0015] Step 5: By solving the rolling optimization of the comprehensive cost function, the optimal three-axis thrust acceleration and yaw rate are obtained, which is the control law for active safety roll guidance of the aircraft.

[0016] It also includes step six, which outputs the control law generated in step five to the aircraft's underlying control system. Based on the rolling time domain control strategy, it only executes the control quantity at the current moment and repeats the above steps at the next moment until the aircraft achieves safe obstacle avoidance, accurate tracking and active perception coordinated control in a dynamic and uncertain environment.

[0017] Furthermore, the dynamic model described in step one is expressed as follows:

[0018]

[0019] in, Indicates the location of the aircraft. Indicates the speed of the aircraft. Indicates the acceleration of the aircraft. Indicates the yaw angle of the aircraft. This indicates the yaw rate of the aircraft; and Constrained by physical mobility, satisfy and .

[0020] Furthermore, the implementation method for step two is as follows:

[0021] Collect historical real trajectory points of the target obstacle, and select valid historical trajectory points for inclusion into the training set sequence using the spatiotemporal information entropy criterion. The selection criteria are as follows:

[0022]

[0023] in, Location of the obstacle. The number of valid historical trajectory points. The current location of the obstacle. The locations of obstacles to be selected in the training set sequence; For the current moment, The time corresponding to the trajectory point in the historical buffer. and These are the spatial displacement and time decay weights, respectively. The attenuation rate, The information entropy threshold;

[0024] For three-dimensional spatial coordinates Construct three independent Gaussian process regression models, using the rational quadratic kernel function as the covariance function:

[0025]

[0026] in, For signal variance, For length scale, For shape parameters, For noise variance; , These are the time nodes corresponding to the i-th and j-th historical valid trajectory points in the training set sequence, respectively. This refers to the Kronecker delta function.

[0027] Using training set sequences The hyperparameter set of the kernel function is optimized online by minimizing the negative log-marginal likelihood function (NLML). ;

[0028] For the first in the future prediction time domain Each prediction step time node ( The mean vector of the predicted position of the obstacle at that moment is obtained through the following Gaussian process posterior inference formula. :

[0029]

[0030] and the variance of the positional uncertainty of obstacles in various dimensions :

[0031]

[0032] in, To train the input time nodes covariance matrix For test nodes Between training nodes Covariance vector It is the identity matrix. This refers to the column vectors of position observations in the corresponding dimension of the training set. The prior covariance at the test node;

[0033] Furthermore, the unified dynamic risk field model described in step three is as follows:

[0034]

[0035] Among them, the Prediction step, spacecraft position Mean of predicted obstacle location The spatial Euclidean distance between them is Obstacle prediction speed With the speed of the aircraft The relative velocity vector is ; It is a distance-sensitive weight constant. It is a speed-sensitive weighting constant. It is the preset maximum reference relative velocity;

[0036] Overall uncertainty variance .

[0037] Furthermore, the safety boundary constraint barrier function with dynamic adaptive expansion mechanism described in step four is expressed as follows:

[0038]

[0039]

[0040] in, The safety boundary constraint weight constant; For the first Predict the constraint state function of the aircraft relative to the dynamic safety boundary of the obstacle during the step; , , The first The components of the predicted position of the aircraft on the three-dimensional spatial axes during the prediction step; , , The first Predicting the mean vector of the future predicted position of dynamic obstacles during the step Components on the axes in three-dimensional space; , , These are the axis length factors of the three-dimensional safety boundary ellipsoid after dynamic adaptive dilation based on the positional uncertainty variance described in step two; where... These are the inherent safety radius reference values ​​for the horizontal and vertical directions, respectively. This is the inflation coefficient for risk uncertainty;

[0041] The trajectory contour error cost function based on risk adaptive adjustment is expressed as follows:

[0042]

[0043]

[0044] in, The baseline weight is constant. The coordinates of the starting waypoint of the global reference path. The unit direction vector of the global reference line. This is a risk-adaptive exponential decay factor.

[0045] Considering the cost function of the active perception value field for line-of-sight occlusion, it is expressed as follows:

[0046]

[0047]

[0048] in, It is the orientation vector of the aircraft's nose camera. It is the gaze vector pointing to the predicted mean of obstacles. To perceive the weight constant, The spatial distance attenuation factor for the perceived value field.

[0049] Heading alignment objective cost function , is represented as ,

[0050]

[0051] in, A vector pointing to the target. This is the heading weight constant.

[0052] Furthermore, the process of constructing the nonlinear model predictive controller and performing rolling optimization in step five is as follows:

[0053] In the prediction time domain Within this framework, a comprehensive cost function for a nonlinear model predictive controller (NMPC) is established. :

[0054]

[0055] in, The cost function for attracting the target point. To control the input smoothing cost function, The target point location, For the first Predict the aircraft acceleration during the step. For the first Predicting the aircraft's heading angular rate at the time of the step. The attraction weight constant for the target point, To control the input smoothing weight constant.

[0056] Furthermore, based on the comprehensive cost function, the overall optimization objective and constraints of nonlinear model predictive control are constructed and summarized, and their standard mathematical form is expressed as follows:

[0057]

[0058] in, To predict the control sequence to be optimized in the time domain, The state transition function is obtained by discretizing the continuous kinematics and dynamics equations over time. The actual state measurement value of the aircraft during the current control cycle;

[0059] The above minimization problem with physical limits and dynamic boundary constraints is solved using a numerical optimization algorithm to obtain the optimal control sequence in the prediction time domain. Extract the optimal triaxial control acceleration from the first step of the sequence. and yaw rate As the current control output.

[0060] Furthermore, the coordinated control process of the aircraft in step six is ​​as follows:

[0061] NMPC employs a rolling optimization mechanism, which, within each control cycle: ① acquires the current state of the aircraft. The system performs the following steps: ① Infers updated obstacle trajectory information through Gaussian processes; ② Generates a safe optimal control sequence by solving the controller model from step five; ③ Outputs only the optimal three-axis control acceleration and yaw rate commands from the first cycle of the sequence to the aircraft's underlying control system to drive the aircraft's motion; ④ When the next control cycle arrives, the system state slides backward by one step, repeating the above prediction and optimization process; This closed-loop rolling feedback compensates for Gaussian process prediction errors and environmental disturbances, enabling safe obstacle avoidance, accurate tracking, and active perception-coordinated control of swaying obstacles in dynamic and uncertain environments.

[0062] Beneficial effects:

[0063] 1. High-confidence uncertainty prediction: The present invention discloses an active safety roll guidance method for aircraft based on Gaussian process, which adopts Gaussian process regression (GPR) with spatiotemporal information entropy adaptive sampling. It can optimize hyperparameters online with limited computing power, accurately output the mean and position variance of the future trajectory of obstacles, and improve the accuracy and efficiency of active safety roll guidance for aircraft.

[0064] 2. Unified Dynamic Risk Field Adaptive Obstacle Avoidance: Breaking the limitations of a fixed safety radius. This invention discloses an active safety roll guidance method for aircraft based on Gaussian processes. The risk field is exponentially coupled with spatial distance, relative velocity, and prediction variance. When facing obstacles approaching at high speed or with high uncertainty, this invention can automatically expand a safety collision avoidance ellipsoid to provide sufficient safety margin.

[0065] 3. Intelligent Trade-off Between Active Perception and Tracking: This invention discloses an active safety roll guidance method for aircraft based on Gaussian processes, which introduces a risk exponential decay mechanism into the NMPC cost function. In highly dangerous situations, the system automatically reduces the tracking weight of the global path profile, prioritizing obstacle avoidance; simultaneously, it actively drives the nose to align with obstacles, ensuring that the perception line of sight is not lost, forming a biomimetic intelligent behavior of "the more dangerous, the more attention," thus achieving safe roll guidance. Attached Figure Description

[0066] Figure 1 The overall flowchart of an active safety roll guidance method for aircraft based on Gaussian processes provided by the present invention;

[0067] Figure 2 This is a panoramic view of the active safety flight trajectory and the dynamic evolution of the sensing cone in a specific embodiment of the present invention. Detailed Implementation

[0068] To better illustrate the purpose and advantages of the present invention, the invention will be further described below in conjunction with the accompanying drawings and examples.

[0069] Example 1:

[0070] like Figure 1 As shown in the figure, this embodiment discloses an active safety roll guidance method for aircraft based on Gaussian processes. The nonlinear model predictive controller has a prediction time-domain step count N=3 and a control cycle step size of 0.4 s. The specific implementation steps are as follows:

[0071] This embodiment discloses an active safety roll guidance method for aircraft based on Gaussian processes, and the specific implementation steps are as follows:

[0072] Step 1: Establish the dynamic model of the aircraft.

[0073] Step 2: During the flight of the aircraft, establish a training set sequence of historical effective trajectory points based on adaptive sampling of spatiotemporal information entropy; use the training set to optimize the set of hyperparameters of the kernel function through online maximum likelihood estimation; calculate the mean vector of the predicted position and its position uncertainty variance in the future prediction time domain of the dynamic obstacle using the Gaussian process posterior inference formula.

[0074] Step 3: Based on the dynamic model in Step 1 and the predicted position mean vector and position uncertainty variance output in Step 2, establish a unified dynamic risk field that combines spatial Euclidean distance, relative velocity, and comprehensive uncertainty variance.

[0075] Step 4: Based on the unified dynamic risk field described in Step 3, establish a safety boundary constraint obstacle function with a dynamic adaptive expansion mechanism, a trajectory contour error cost function based on risk adaptive adjustment, an active perception value field cost function considering line-of-sight occlusion, and a heading alignment target cost function; introduce a target point attraction cost function and a control input smoothing cost function to construct a comprehensive cost function.

[0076] Step 5: By solving the rolling optimization of the comprehensive cost function, the optimal three-axis thrust acceleration and yaw rate are obtained, which is the control law for active safety roll guidance of the aircraft.

[0077] Step 6: Output the control law generated in Step 5 to the aircraft's underlying control system. Based on the rolling time domain control strategy, only execute the control quantity at the current moment. Repeat the above steps at the next moment until the aircraft achieves safe obstacle avoidance, accurate tracking, and active perception and coordinated control in a dynamic and uncertain environment.

[0078] The dynamic model described in step one is expressed as follows:

[0079]

[0080] in, Indicates the location of the aircraft. Indicates the speed of the aircraft. Indicates the acceleration of the aircraft. Indicates the yaw angle of the aircraft. This indicates the yaw rate of the aircraft; and Constrained by physical mobility, satisfy and In this embodiment, it is set , .

[0081] The implementation method for step two is as follows:

[0082] Collect historical real trajectory points of the target obstacle, and select valid historical trajectory points for inclusion into the training set sequence using the spatiotemporal information entropy criterion. The selection criteria are as follows:

[0083]

[0084] in, Location of the obstacle. The number of valid historical trajectory points. The current location of the obstacle. The locations of obstacles to be selected in the training set sequence; For the current moment, The time corresponding to the trajectory point in the historical buffer. and These are the spatial displacement and time decay weights, respectively. The attenuation rate, This is the information entropy threshold; in this embodiment, it is set as follows: , , , .

[0085] For three-dimensional spatial coordinates Construct three independent Gaussian process regression models, using the rational quadratic kernel function as the covariance function:

[0086]

[0087] in, For signal variance, For length scale, For shape parameters, For noise variance; , These are the time nodes corresponding to the i-th and j-th historical valid trajectory points in the training set sequence, respectively. This refers to the Kronecker delta function.

[0088] Using training set sequences The hyperparameter set of the kernel function is optimized online by minimizing the negative log-marginal likelihood function (NLML). ;

[0089] For the first in the future prediction time domain Each prediction step time node ( The mean vector of the predicted position of the obstacle at that moment is obtained through the following Gaussian process posterior inference formula. :

[0090]

[0091] and the variance of the positional uncertainty of obstacles in various dimensions :

[0092]

[0093] in, To train the input time nodes covariance matrix For test nodes Between training nodes Covariance vector It is the identity matrix. This refers to the column vectors of position observations in the corresponding dimension of the training set. The prior covariance at the test node;

[0094] The unified dynamic risk field model described in step three is as follows:

[0095]

[0096] Among them, the Prediction step, spacecraft position Mean of predicted obstacle location The spatial Euclidean distance between them is Obstacle prediction speed With the speed of the aircraft The relative velocity vector is ; It is a distance-sensitive weight constant. It is a speed-sensitive weighting constant. It is the preset maximum reference relative velocity;

[0097] Overall uncertainty variance In this embodiment, we take , , .

[0098] The safety boundary constraint barrier function with dynamic adaptive expansion mechanism described in step four is expressed as follows:

[0099]

[0100]

[0101] in, The safety boundary constraint weight constant; For the first Predict the constraint state function of the aircraft relative to the dynamic safety boundary of the obstacle during the step; , , The first The components of the predicted position of the aircraft on the three-dimensional spatial axes during the prediction step; , , The first Predicting the mean vector of the future predicted position of dynamic obstacles during the step Components on the axes in three-dimensional space; , , These are the axis length factors of the three-dimensional safety boundary ellipsoid after dynamic adaptive dilation based on the positional uncertainty variance described in step two; where... These are the inherent safety radius reference values ​​for the horizontal and vertical directions, respectively. This is the risk uncertainty inflation coefficient; in this embodiment, it is set as follows: ;

[0102] The trajectory contour error cost function based on risk adaptive adjustment is expressed as follows:

[0103]

[0104]

[0105] in, The baseline weight is constant. The coordinates of the starting waypoint of the global reference path. The unit direction vector of the global reference line. This is the risk adaptive exponential decay factor; in this embodiment, it is set as follows: , .

[0106] Considering the cost function of the active perception value field for line-of-sight occlusion, it is expressed as follows:

[0107]

[0108]

[0109] in, It is the orientation vector of the aircraft's nose camera. It is the gaze vector pointing to the predicted mean of obstacles. To perceive the weight constant, The spatial distance attenuation factor of the perceived value field; in this embodiment, we take... , .

[0110] Heading alignment objective cost function , is represented as ,

[0111]

[0112] in, A vector pointing to the target. Here is the heading weight constant; in this embodiment, it is taken as... ;

[0113] The process of constructing the nonlinear model predictive controller and performing rolling optimization in step five is as follows:

[0114] In the prediction time domain Within this framework, a comprehensive cost function for a nonlinear model predictive controller (NMPC) is established. :

[0115]

[0116] in, The cost function for attracting the target point. To control the input smoothing cost function, The target point location, For the first Predict the aircraft acceleration during the step. For the first Predicting the aircraft's heading angular rate at the time of the step. The attraction weight constant for the target point, To control the input smoothing weight constant, this embodiment uses weights... , .

[0117] Based on the comprehensive cost function, the overall optimization objective and constraints of nonlinear model predictive control are constructed and summarized, and their standard mathematical form is as follows:

[0118]

[0119] in, To predict the control sequence to be optimized in the time domain, The state transition function is obtained by time discretization of the continuous dynamic equation as described in claim 2. The actual state measurement value of the aircraft during the current control cycle;

[0120] The above minimization problem with physical limits and dynamic boundary constraints is solved using a numerical optimization algorithm to obtain the optimal control sequence in the prediction time domain. Extract the optimal triaxial control acceleration from the first step of the sequence. and yaw rate As the current control output.

[0121] The coordinated control process of the aircraft in step six is ​​as follows:

[0122] NMPC employs a rolling optimization mechanism, which, within each control cycle: ① acquires the current state of the aircraft. The process involves: ① inferring updated obstacle trajectory information through Gaussian processes; ② generating a safe optimal control sequence by solving the controller model from step five; ③ outputting the optimal three-axis control acceleration and yaw rate commands for only the first cycle of the sequence to the aircraft's underlying control system to drive the aircraft's motion; ④ when the next control cycle arrives, the system state slides backward by one step, repeating the above prediction and optimization process; This closed-loop rolling feedback compensates for Gaussian process prediction errors and environmental disturbances, achieving safe obstacle avoidance, precise tracking, and active perception-coordinated control of swaying obstacles in dynamic and uncertain environments. The final implementation effect is as follows: Figure 2 As shown.

[0123] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for active safety roll guidance of an aircraft based on Gaussian processes, characterized in that: Includes the following steps, Step 1: Establish the dynamic model of the aircraft; Step 2: During the flight of the aircraft, establish a training set sequence of historical effective trajectory points based on adaptive sampling of spatiotemporal information entropy; use the training set to optimize the set of hyperparameters of the kernel function through online maximum likelihood estimation; and calculate the mean vector of the predicted position and its position uncertainty variance in the future prediction time domain of the dynamic obstacle using the Gaussian process posterior inference formula. Step 3: Based on the dynamic model, the predicted position mean vector and the position uncertainty variance, establish a unified dynamic risk field that combines spatial Euclidean distance, relative velocity and comprehensive uncertainty variance; Step 4: Based on the unified dynamic risk field, establish the safety boundary constraint obstacle function with dynamic adaptive expansion mechanism, the trajectory contour error cost function based on risk adaptive adjustment, the active perception value field cost function considering line-of-sight occlusion, and the heading alignment target cost function; introduce the target point attraction cost function and the control input smoothing cost function to construct the comprehensive cost function. Step 5: By solving the rolling optimization of the comprehensive cost function, the optimal three-axis thrust acceleration and yaw rate are obtained, which is the control law for active safety roll guidance of the aircraft.

2. The method as described in claim 1, characterized in that: It also includes step six, which outputs the control law generated in step five to the aircraft's underlying control system. Based on the rolling time domain control strategy, only the control quantity at the current moment is executed, and the above steps are repeated at the next moment until the aircraft achieves safe obstacle avoidance, accurate tracking and active perception coordinated control in a dynamic and uncertain environment.

3. The method as described in claim 1 or 2, characterized in that: The dynamic model described in step one is expressed as follows: in, Indicates the location of the aircraft. Indicates the speed of the aircraft. Indicates the acceleration of the aircraft. Indicates the yaw angle of the aircraft. This indicates the yaw rate of the aircraft; and Constrained by physical mobility, satisfy and .

4. The method as described in claim 3, characterized in that: The implementation method for step two is as follows: Collect historical real trajectory points of the target obstacle, and filter valid historical trajectory points into the training set sequence using the spatiotemporal information entropy criterion. The selection criteria are as follows: in, Location of the obstacle. The number of valid historical trajectory points. The current location of the obstacle. The locations of obstacles to be selected in the training set sequence; For the current moment, The time corresponding to the trajectory point in the historical buffer. and These are the spatial displacement and time decay weights, respectively. The attenuation rate, The information entropy threshold; For three-dimensional spatial coordinates Construct three independent Gaussian process regression models, using the rational quadratic kernel function as the covariance function: in, For signal variance, For length scale, For shape parameters, For noise variance; , These are the time nodes corresponding to the i-th and j-th historical valid trajectory points in the training set sequence, respectively. For the Kronecker delta function; Using training set sequences The hyperparameter set of the kernel function is optimized online by minimizing the negative log-marginal likelihood function (NLML). ; For the first in the future prediction time domain Each prediction step time node ( The mean vector of the predicted position of the obstacle at that moment is obtained through the following Gaussian process posterior inference formula. : and the variance of the positional uncertainty of obstacles in various dimensions : in, To train the input time nodes covariance matrix For test nodes Between training nodes Covariance vector It is the identity matrix. For the corresponding dimension of the position observation column vector in the training set; Let be the prior covariance at the test node.

5. The method as described in claim 4, characterized in that: the unified dynamic risk field model in step three is, in, No. Prediction step, spacecraft position Mean of predicted obstacle location The spatial Euclidean distance between them is Obstacle prediction speed With the speed of the aircraft The relative velocity vector is ; It is a distance-sensitive weight constant. It is a speed-sensitive weighting constant. It is the preset maximum reference relative velocity; Overall uncertainty variance .

6. The method as described in claim 5, characterized in that: The safety boundary constraint barrier function with dynamic adaptive expansion mechanism described in step four is expressed as follows: in, The safety boundary constraint weight constant; For the first Predict the constraint state function of the aircraft relative to the dynamic safety boundary of the obstacle during the step; , , The first The components of the predicted position of the aircraft on the three-dimensional spatial axes during the prediction step; , , The first Predicting the mean vector of the future predicted position of dynamic obstacles during the step Components on the axes in three-dimensional space; , , These are the axis length factors of the three-dimensional safety boundary ellipsoid after dynamic adaptive dilation based on the positional uncertainty variance described in step two; where... These are the inherent safety radius reference values ​​for the horizontal and vertical directions, respectively. This is the inflation coefficient for risk uncertainty; The trajectory contour error cost function based on risk adaptive adjustment is expressed as follows: in, The baseline weight is constant. The coordinates of the starting waypoint of the global reference path. The unit direction vector of the global reference line. This is a risk-adaptive exponential decay factor. Considering the cost function of the active perception value field for line-of-sight occlusion, it is expressed as follows: in, It is the orientation vector of the aircraft's nose camera. It is the gaze vector pointing to the predicted mean of obstacles. To perceive the weight constant, The spatial distance attenuation factor for the perceived value field; Heading alignment objective cost function , is represented as , in, A vector pointing to the target. This is the heading weight constant.

7. The active safety roll guidance method for aircraft based on Gaussian processes as described in claim 6, characterized in that: The process of constructing the nonlinear model predictive controller and performing rolling optimization in step five is as follows: In the prediction time domain Within this framework, a comprehensive cost function for a nonlinear model predictive controller (NMPC) is established. : in, The cost function for attracting the target point. To control the input smoothing cost function, The target point location, For the first Predict the aircraft acceleration during the step. For the first Predicting the aircraft's heading angular rate at the time of the step. The attraction weight constant for the target point, To control the input smoothing weight constant.

8. The active safety roll guidance method for aircraft based on Gaussian processes as described in claim 7, characterized in that: Based on the comprehensive cost function, the overall optimization objective and constraints of nonlinear model predictive control are constructed and summarized, and their standard form is expressed as follows: in, To predict the control sequence to be optimized in the time domain, The state transition function is obtained by time discretization of the continuous kinematics and dynamics equations as described in claim 2. The actual state measurement value of the aircraft during the current control cycle; The above minimization problem with physical limits and dynamic boundary constraints is solved using a numerical optimization algorithm to obtain the optimal control sequence in the prediction time domain. Extract the optimal triaxial control acceleration from the first step of the sequence. and yaw rate As the current control output.

9. The active safety roll guidance method for aircraft based on Gaussian processes as described in claim 8, characterized in that: The coordinated control process of the aircraft in step six is ​​as follows: NMPC employs a rolling optimization mechanism, which, within each control cycle: ① acquires the current state of the aircraft. The system performs the following steps: ① Infers updated obstacle trajectory information through Gaussian processes; ② Generates a safe optimal control sequence by solving the controller model from step five; ③ Outputs only the optimal three-axis control acceleration and yaw rate commands from the first cycle of the sequence to the aircraft's underlying control system to drive the aircraft's motion; ④ When the next control cycle arrives, the system state slides backward by one step, repeating the above prediction and optimization process; This closed-loop rolling feedback compensates for Gaussian process prediction errors and environmental disturbances, enabling safe obstacle avoidance, accurate tracking, and active perception-coordinated control of swaying obstacles in dynamic and uncertain environments.

Images