Unmanned aerial vehicle path planning method

Abstract

The invention discloses an unmanned aerial vehicle path planning method, which comprises the following steps: carrying out environment data acquisition and state modeling, and generating an initial task data set containing starting point distribution, end point distribution, a reference random process and original and real-time environment data; modeling and inferring by using a Gaussian process potential force model, and outputting a disturbance force prediction result containing a mean value and uncertainty; on the basis of the starting point, the terminal point, the reference random process and the disturbance force, constructing SRoudinger bridge distribution evolution and multi-objective optimization to obtain an optimal state distribution track; according to the uncertainty and the battery capacity, self-adaptively adjusting the regularization coefficient of the SRosine-Bridge entropy, and updating and optimizing the weight; and mapping the optimal state distribution track into a control input sequence, and triggering re-optimization when the environment suddenly changes. According to the method, Gaussian process potential force modeling and SRoudinger bridge optimization are fused to improve path precision, energy efficiency and environmental adaptability.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control technology, and more particularly to a UAV path planning method. Background Technology

[0002] Currently, UAV path planning technology plays a crucial role in autonomous flight in complex environments. However, existing technologies still have significant shortcomings in environmental perception and path optimization. Traditional methods often rely on data from a single sensor or a limited number of sensors, resulting in incomplete environmental information acquisition. This is especially true when influenced by dynamic factors such as wind speed and direction, turbulence intensity, and weather changes, leading to insufficient path planning accuracy. Furthermore, the lag in environmental data updates hinders rapid responses to sudden environmental changes, impacting flight safety and energy consumption control.

[0003] In terms of path planning model construction, existing technologies generally employ simplified dynamic models and single optimization objectives, lacking a comprehensive consideration of the balance between energy consumption and safety for UAVs under different mission types. These methods typically ignore the spatiotemporal correlation modeling of environmental disturbances, making it difficult to accurately predict disturbances that cannot be directly measured, thus affecting the stability of the planning results. Furthermore, most algorithms lack a dynamic adjustment mechanism between the reference trajectory and the actual flight trajectory, making the path optimization results prone to failure once environmental conditions change during flight.

[0004] Existing UAV path planning methods suffer from problems such as a single optimization objective, insufficient real-time performance, and poor adaptability to highly uncertain environments. They fail to effectively balance path smoothness, energy efficiency, and safety, thus limiting the autonomous operation capabilities of UAVs in complex environments. To address these issues, there is an urgent need for a path planning method that can fully utilize multi-source environmental information, accurately model environmental disturbances, and possess dynamic optimization capabilities to improve the autonomous flight performance and mission reliability of UAVs in complex and dynamic environments.

[0005] Therefore, how to provide a method for UAV path planning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a path planning method for unmanned aerial vehicles (UAVs). This invention integrates multi-source sensor data with a Gaussian process latent force model to accurately predict environmental disturbances. It combines a Scholdinger bridge distribution evolution model with a multi-objective optimization strategy to achieve a smooth transition from the starting state distribution to the ending state distribution and has dynamic adaptive adjustment capabilities. In complex dynamic environments, it can balance path smoothness, energy efficiency, and safety, thereby improving the reliability of autonomous flight and mission execution efficiency of UAVs.

[0007] A method for unmanned aerial vehicle (UAV) path planning according to an embodiment of the present invention includes the following steps:

[0008] Perform environmental data acquisition and state modeling, collect and fuse multi-source sensor data of UAV to obtain probabilistic representations of position and velocity at takeoff time. The probabilistic representations include state estimates and their confidence intervals, which are used to generate the starting state distribution. Collect the ending state distribution of the target area, wind speed, wind direction and turbulence intensity, temperature and humidity, environmental obstacle information obtained by meteorological sensors and visual sensors, and real-time environmental data during flight. Construct an initial task dataset containing the starting state distribution, ending state distribution, reference stochastic process, raw environmental data and real-time environmental data.

[0009] Based on the Gaussian process potential force model, the model is used to model and infer the original environmental data and real-time environmental data in the initial task dataset, predict the environmental disturbance force that cannot be directly measured. The model uses historical flight data and real-time environmental data to learn the spatiotemporal correlation of the disturbance force and generate the disturbance force prediction results including mean estimation and uncertainty.

[0010] Based on the starting state distribution, ending state distribution, reference stochastic process, and perturbation force prediction results, a Scholdinger bridge distribution evolution model is constructed to achieve a smooth transition from the starting state distribution to the ending state distribution while minimizing deviation from the reference stochastic process. At the same time, energy consumption cost and collision risk cost are introduced to form a multi-objective optimization model, and the optimal state distribution trajectory is output.

[0011] The Scholdinger bridge entropy regularization coefficient is adaptively adjusted. The Scholdinger bridge entropy regularization coefficient adjusts the constraint strength of the distribution evolution process on the reference random process, controls the smoothness of the state distribution trajectory and the energy consumption trade-off. The regularization coefficient is dynamically adjusted according to the uncertainty in the disturbance prediction results and the remaining battery capacity of the UAV. When the uncertainty increases or the remaining battery capacity decreases, the coefficient is increased to increase path smoothness and energy saving.

[0012] The optimal state distribution trajectory is mapped to the control input sequence of the UAV. When the uncertainty of the disturbance force prediction exceeds the set threshold, the Scholdinger bridge optimization is re-triggered based on the real-time update result of the Gaussian process potential force model during flight to cope with sudden environmental changes.

[0013] Optionally, the construction of the initial task dataset includes:

[0014] The system collects multi-source sensor data from the UAV, including observation data from position sensors, velocity sensors, inertial measurement units, meteorological sensors, and vision sensors. It synchronizes the data with time and aligns the timestamps according to a unified time base. The collected multi-source sensor data is fused, and the state estimates and confidence intervals of position and velocity are calculated using a multi-sensor fusion algorithm to generate the starting state distribution. The multi-sensor fusion algorithm is a mature method in the existing technology for fusing multi-source sensor observation data to generate state estimates and confidence intervals, including extended Kalman filtering, unscented Kalman filtering, particle filtering, and factor graph-based optimization methods.

[0015] The system collects the endpoint state distribution and raw environmental data of the target area. The raw environmental data includes wind speed and direction, turbulence intensity, temperature and humidity and environmental obstacle information of the flight area. Real-time environmental data is also collected during the flight. The raw environmental data and real-time environmental data are analyzed and calculated to generate environmental disturbance parameters.

[0016] A reference stochastic process is established by combining the flight dynamics model of the UAV. By mapping environmental disturbance parameters to state transition probability distribution, the reference stochastic process is generated. The state transition probability distribution characterizes the state change law of the UAV under the influence of the natural environment under the condition of no control input. The flight dynamics model of the UAV is a publicly available UAV dynamics modeling method, including translational and rotational motion equations.

[0017] The starting state distribution, ending state distribution, reference stochastic process, raw environmental data, and real-time environmental data are encapsulated into an initial task dataset.

[0018] Optionally, the generation of the disturbance prediction result includes:

[0019] Raw environmental data and real-time environmental data are extracted from the initial task dataset to construct a disturbance force prediction sample set. The sample inputs include wind speed and direction, turbulence intensity, temperature and humidity, and environmental obstacle information. The sample output is the environmental disturbance force.

[0020] A Gaussian process latent force model is established, a spatiotemporal kernel function is selected, and kernel parameters and noise parameters are set. The spatiotemporal kernel function establishes the correlation between spatial location, time, and meteorological variables. The kernel parameters are initialized from historical flight data. The noise parameters are used to characterize the measurement error of multi-source sensor observation data and affect the uncertainty calculation of the disturbance force prediction results. The spatiotemporal kernel function can take the following form:

[0021]

[0022] Where, x i ,x jThe vectors representing the three-dimensional spatial positions of the i-th and j-th samples, in meters, are derived from the state estimates obtained by fusing the UAV position sensor and inertial measurement unit. i ,t j This represents the timestamps of the i-th and j-th samples, in seconds, derived from the sampling time after the task time base is synchronized. i ,z j Let τ represent the meteorological variable vectors for the i-th and j-th samples, including wind speed, wind direction, temperature, and humidity, derived from the fusion data of airborne meteorological sensors and visual sensors. i ,τ j This represents the turbulence intensity corresponding to the i-th and j-th samples, expressed as a dimensionless ratio, derived from the normalized value of the turbulence energy dissipation rate calculated by the meteorological sensor. σ f Let l be the variance of the kernel function signal, representing the prior scale of the Gaussian process for the magnitude of the perturbation force change, initialized to the variance of the perturbation force prediction sample set. s The spatial length scale represents the characteristic length at which the correlation of the disturbance force decays with spatial distance. It is determined based on the half-decay distance of the autocorrelation function of the disturbance force in historical flight trajectories, and the unit is meters (l). t The time scale represents the characteristic time at which the correlation of the disturbance force decays with time intervals. It is determined based on the time interval at which the time autocorrelation function of the disturbance force decays to 1 / e, and the unit is seconds. z The length scale of the meteorological variable represents the characteristic scale at which the correlation of disturbance forces decays with differences in meteorological conditions. It is calculated based on the variance ratio of the changes in disturbance forces under different meteorological conditions, and the unit is the normalized scale of the meteorological variable. α is the turbulence modulation coefficient, which represents the amplification effect of turbulence intensity on the correlation of disturbance forces. Its value is obtained by fitting the Pearson correlation coefficient between turbulence intensity and disturbance forces in historical flight data, and its range is between 0 and 1.

[0023] The selection of kernel parameters is based on the following criteria: spatial length scale l s The time scale is determined based on the sensitivity of perturbation forces to changes in spatial position in historical flight trajectories. t The length scale of meteorological variables is determined based on the time interval when the time autocorrelation function of the disturbance decays to 1 / e. z The signal variance σ is calculated based on the variance ratio of the disturbance force changes under different meteorological conditions. f Based on the variance initialization of the disturbance force prediction sample set, the turbulence modulation coefficient α is obtained by fitting the correlation coefficient between turbulence intensity and disturbance force in historical data;

[0024] The Gaussian process potential force model is trained offline based on historical flight data, and online parameter updates are performed based on real-time environmental data during flight to complete spatiotemporal correlation learning.

[0025] The raw environmental data and real-time environmental data are input into the Gaussian process potential force model to generate disturbance force prediction results, which include mean estimation and uncertainty.

[0026] The perturbation prediction results are indexed with the starting state distribution, ending state distribution and reference stochastic process in the initial task dataset, and the perturbation prediction results are output.

[0027] Optionally, the generation of the optimal state distribution trajectory includes:

[0028] Based on the starting state distribution, ending state distribution, reference stochastic process, and disturbance prediction results, the Scholdinger bridge distribution evolution model is initialized. The starting state distribution and ending state distribution are used as boundary conditions, the reference stochastic process is used as the state evolution benchmark under no control input conditions, and the disturbance prediction results are introduced into the evolution process as external disturbance terms.

[0029] A multi-objective optimization model is constructed. The objective function of the multi-objective optimization model includes energy consumption cost and collision risk cost. The energy consumption cost is calculated by the cumulative thrust input and flight time, and the collision risk cost is calculated by the minimum distance between the optimal state distribution trajectory and the safety boundary of environmental obstacles. The energy consumption cost and collision risk cost are combined in a weighted manner to form the total objective function. The weight coefficients are dynamically adjusted according to the UAV mission type, remaining battery capacity and environmental congestion.

[0030] The Scholdinger bridge distribution evolution is iteratively solved based on a multi-objective optimization model. During the iteration process, the total objective function and the Scholdinger bridge constraint terms are minimized simultaneously to achieve a balance between smoothness, energy efficiency and safety of the state distribution trajectory. The optimal state distribution trajectory is output after convergence.

[0031] Optionally, the adaptive adjustment of the Scholdinger bridge entropy regularization coefficient includes:

[0032] Uncertainty is extracted from the disturbance force prediction results and, together with the remaining battery capacity of the UAV, constitutes the adjustment input. Normalization and denoising are then performed to generate the adjustment input vector.

[0033] The Scholdinger bridge entropy regularization coefficient is updated based on the adjustment input vector within a preset upper and lower bound. The update step size and monotonic mapping rule are set to generate the current Scholdinger bridge entropy regularization coefficient.

[0034] The current Scholdinger bridge entropy regularization coefficient is applied to the Scholdinger bridge distribution evolution model to update the weights of the objective terms related to the reference stochastic process in the multi-objective optimization model. The weights of energy consumption cost and collision risk cost are synchronously scheduled according to the task configuration to generate the updated optimization problem.

[0035] During flight, a threshold triggering and hysteresis mechanism is executed according to the control cycle. When the uncertainty of the disturbance prediction result exceeds the threshold or the remaining battery capacity of the UAV is lower than the threshold, the Scholdinger bridge entropy regularization coefficient is increased. When the uncertainty and the remaining battery capacity of the UAV enter the hysteresis interval, the Scholdinger bridge entropy regularization coefficient is decreased.

[0036] Optionally, the process of mapping the optimal state distribution trajectory to the UAV control input sequence includes:

[0037] Receive discrete state points in the optimal state distribution trajectory, calculate the attitude adjustment amount and thrust requirement corresponding to each state point according to the UAV's flight dynamics model, and convert the calculation results into a set of control input commands;

[0038] According to the control cycle, the control input command set is sent to the UAV flight control system to execute the corresponding attitude adjustment and thrust control, so that the UAV flies according to the optimal state distribution trajectory.

[0039] During flight, the system receives real-time updates of the disturbance force prediction results from the Gaussian process potential force model. When the uncertainty of the disturbance force prediction results exceeds the threshold or the remaining battery capacity of the UAV is lower than the threshold, the Scholdinger bridge entropy regularization coefficient is increased. When the uncertainty and the remaining battery capacity of the UAV enter the hysteresis interval, the Scholdinger bridge entropy regularization coefficient is decreased, triggering the re-optimization process of the Scholdinger bridge distribution evolution model and recalculating the optimal state distribution trajectory.

[0040] The optimal state distribution trajectory obtained through re-optimization is mapped to a new set of control input instructions, which replaces the currently executed instructions within the control cycle, enabling a rapid response to sudden environmental changes.

[0041] The beneficial effects of this invention are:

[0042] First, by fusing multi-source sensor data with a Gaussian process latent force model, this invention achieves high-precision prediction of environmental disturbance forces, which can effectively reduce trajectory deviations caused by factors such as meteorological changes and turbulence.

[0043] Secondly, by combining the Scholdinger bridge distribution evolution model with a multi-objective optimization strategy, the energy consumption cost and collision risk cost can be taken into account simultaneously during the path planning process, so that the UAV can maintain the smoothness and safety of the flight path in complex environments.

[0044] Furthermore, the introduction of a dynamic adaptive adjustment mechanism with Scholdinger bridge entropy regularization coefficient enables the system to adjust its path planning strategy in real time according to changes in environmental uncertainty and remaining battery capacity. This allows the system to maintain a high mission completion rate and energy efficiency in the event of sudden environmental changes, significantly improving the autonomous flight performance and robustness of the UAV in dynamic and complex mission scenarios. Attached Figure Description

[0045] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0046] Figure 1 This is an overall flowchart of a UAV path planning method proposed in this invention;

[0047] Figure 2 This is a schematic diagram of the structure for generating perturbation force prediction results based on the Gaussian process latent force model proposed in this invention.

[0048] Figure 3 This is a schematic diagram of the process for generating the optimal state distribution trajectory by combining the Scholdinger bridge distribution evolution model and the multi-objective optimization model proposed in this invention. Detailed Implementation

[0049] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0050] refer to Figure 1-3 A method for unmanned aerial vehicle (UAV) path planning includes the following steps:

[0051] Perform environmental data acquisition and state modeling, collect and fuse multi-source sensor data of UAV to obtain probabilistic representations of position and velocity at takeoff time. The probabilistic representations include state estimates and their confidence intervals, which are used to generate the starting state distribution. Collect the ending state distribution of the target area, wind speed, wind direction and turbulence intensity, temperature and humidity, environmental obstacle information obtained by meteorological sensors and visual sensors, and real-time environmental data during flight. Construct an initial task dataset containing the starting state distribution, ending state distribution, reference stochastic process, raw environmental data and real-time environmental data.

[0052] Based on the Gaussian process potential force model, the model is used to model and infer the original environmental data and real-time environmental data in the initial task dataset, predict the environmental disturbance force that cannot be directly measured. The model uses historical flight data and real-time environmental data to learn the spatiotemporal correlation of the disturbance force and generate the disturbance force prediction results including mean estimation and uncertainty.

[0053] Based on the starting state distribution, ending state distribution, reference stochastic process, and perturbation force prediction results, a Scholdinger bridge distribution evolution model is constructed to achieve a smooth transition from the starting state distribution to the ending state distribution while minimizing deviation from the reference stochastic process. At the same time, energy consumption cost and collision risk cost are introduced to form a multi-objective optimization model, and the optimal state distribution trajectory is output.

[0054] The Scholdinger bridge entropy regularization coefficient is adaptively adjusted. The Scholdinger bridge entropy regularization coefficient adjusts the constraint strength of the distribution evolution process on the reference random process, controls the smoothness of the state distribution trajectory and the energy consumption trade-off. The regularization coefficient is dynamically adjusted according to the uncertainty in the disturbance prediction results and the remaining battery capacity of the UAV. When the uncertainty increases or the remaining battery capacity decreases, the coefficient is increased to increase path smoothness and energy saving.

[0055] The optimal state distribution trajectory is mapped to the control input sequence of the UAV. When the uncertainty of the disturbance force prediction exceeds the set threshold, the Scholdinger bridge optimization is re-triggered based on the real-time update result of the Gaussian process potential force model during flight to cope with sudden environmental changes.

[0056] This invention proposes a path planning method for unmanned aerial vehicles (UAVs). It generates an initial task dataset containing starting state distribution, ending state distribution, a reference stochastic process, raw environmental data, and real-time environmental data through environmental data acquisition and state modeling. The method uses a Gaussian process latent force model to predict disturbance forces, combines a Scholdinger bridge distribution evolution model with multi-objective optimization to achieve smooth transitions in state distribution, and introduces an adaptive adjustment mechanism for the Scholdinger bridge entropy regularization coefficient and real-time optimization control input mapping to achieve high-precision, low-energy-consumption path planning with dynamic environmental adaptability.

[0057] In this embodiment, the construction of the initial task dataset includes:

[0058] The system collects multi-source sensor data from the UAV, including observation data from position sensors, velocity sensors, inertial measurement units, meteorological sensors, and vision sensors. It synchronizes the data with time and aligns the timestamps according to a unified time base. The collected multi-source sensor data is fused, and the state estimates and confidence intervals of position and velocity are calculated using a multi-sensor fusion algorithm to generate the starting state distribution. The multi-sensor fusion algorithm is a mature method in the existing technology for fusing multi-source sensor observation data to generate state estimates and confidence intervals, including extended Kalman filtering, unscented Kalman filtering, particle filtering, and factor graph-based optimization methods.

[0059] The system collects the endpoint state distribution and raw environmental data of the target area. The raw environmental data includes wind speed and direction, turbulence intensity, temperature and humidity and environmental obstacle information of the flight area. Real-time environmental data is also collected during the flight. The raw environmental data and real-time environmental data are analyzed and calculated to generate environmental disturbance parameters.

[0060] A reference stochastic process is established by combining the flight dynamics model of the UAV. By mapping environmental disturbance parameters to state transition probability distribution, the reference stochastic process is generated. The state transition probability distribution characterizes the state change law of the UAV under the influence of the natural environment under the condition of no control input. The flight dynamics model of the UAV is a publicly available UAV dynamics modeling method, including translational and rotational motion equations.

[0061] The starting state distribution, ending state distribution, reference stochastic process, raw environmental data, and real-time environmental data are encapsulated into an initial task dataset.

[0062] The specific process for generating environmental disturbance parameters by analyzing and calculating raw and real-time environmental data is as follows: First, the raw and real-time environmental data are synchronized and timestamp-aligned to an inertial coordinate system, and quality gating and outlier removal are performed. Local interpolation is used for short-term missing data, and high-frequency noise is smoothed. Second, based on the relationship between wind speed and direction and flight altitude and position, the wind field is decomposed into mean wind vector, gust component, and vertical airflow component. The spatial gradient and shear intensity varying with altitude are calculated. Combined with turbulence intensity and its short-time variance, statistics reflecting instability are obtained. Air density is estimated using temperature and humidity to form a density index related to the magnitude of aerodynamic forces. Environmental obstacle information is mapped to a gridded area, and risk weights that may induce backflow and confluence are marked. Finally, an environmental disturbance parameter set is generated according to the control cycle. The set includes mean wind vector, gust variance and correlation, vertical airflow intensity, wind shear and spatial gradient, turbulence intensity statistics, air density estimate, and obstacle-induced flow field indicator, with corresponding time steps and spatial resolutions provided for subsequent modeling input.

[0063] The state transition probability distribution is constructed based on the combined effect of the UAV flight dynamics model and environmental disturbance parameters. First, using the starting state of the current control cycle, including position, velocity, and attitude, as input, the UAV flight dynamics model calculates the state change trend under the influence of only natural forces such as gravity, air resistance, and lift, without applying control input. Second, environmental disturbance parameters are introduced into the dynamic calculation. These parameters include the mean wind vector, vertical airflow component, wind shear intensity, turbulence intensity, air density, and obstacle-induced flow field information. The mean wind vector is used to correct for changes in position and velocity; the vertical airflow component and air density are used to adjust lift and vertical velocity; wind shear and turbulence intensity, as random disturbance sources, reflect uncertainty; and obstacle-induced flow field information is used to increase the disturbance intensity within a specific spatial region. Then, based on the state change results after introducing the disturbance, the expected change and uncertainty range of the state transition are determined. The expected change reflects the average change in position and velocity within a control cycle, while the uncertainty range is jointly determined by factors such as wind shear, turbulence intensity, gust intensity, and obstacle disturbance, and is used to characterize the magnitude and correlation of uncertainty between different state components. Finally, the state transition probability distribution is defined by the expected change and uncertainty range, and dynamically adjusted according to the updated environmental disturbance parameters in each new control cycle. This probability distribution can reflect the natural state evolution characteristics under the current environmental conditions in real time, providing a reliable reference basis for subsequent distribution evolution.

[0064] In this embodiment, the generation of the disturbance prediction result includes:

[0065] Raw environmental data and real-time environmental data are extracted from the initial task dataset to construct a disturbance force prediction sample set. The sample inputs include wind speed and direction, turbulence intensity, temperature and humidity, and environmental obstacle information. The sample output is the environmental disturbance force.

[0066] A Gaussian process latent force model is established, a spatiotemporal kernel function is selected, and kernel parameters and noise parameters are set. The spatiotemporal kernel function establishes the correlation between spatial location, time, and meteorological variables. The kernel parameters are initialized from historical flight data. The noise parameters are used to characterize the measurement error of multi-source sensor observation data and affect the uncertainty calculation of the disturbance force prediction results. The spatiotemporal kernel function can take the following form:

[0067]

[0068] Where, x i ,x j The vectors representing the three-dimensional spatial positions of the i-th and j-th samples, in meters, are derived from the state estimates obtained by fusing the UAV position sensor and inertial measurement unit. i ,t jThis represents the timestamps of the i-th and j-th samples, in seconds, derived from the sampling time after the task time base is synchronized. i ,z j Let τ represent the meteorological variable vectors for the i-th and j-th samples, including wind speed, wind direction, temperature, and humidity, derived from the fusion data of airborne meteorological sensors and visual sensors. i ,τ j This represents the turbulence intensity corresponding to the i-th and j-th samples, expressed as a dimensionless ratio, derived from the normalized value of the turbulence energy dissipation rate calculated by the meteorological sensor. σ f Let l be the variance of the kernel function signal, representing the prior scale of the Gaussian process for the magnitude of the perturbation force change, initialized to the variance of the perturbation force prediction sample set. s The spatial length scale represents the characteristic length at which the correlation of the disturbance force decays with spatial distance. It is determined based on the half-decay distance of the autocorrelation function of the disturbance force in historical flight trajectories, and the unit is meters (l). t The time scale represents the characteristic time at which the correlation of the disturbance force decays with time intervals. It is determined based on the time interval at which the time autocorrelation function of the disturbance force decays to 1 / e, and the unit is seconds. z The length scale of the meteorological variable represents the characteristic scale at which the correlation of disturbance forces decays with differences in meteorological conditions. It is calculated based on the variance ratio of the changes in disturbance forces under different meteorological conditions, and the unit is the normalized scale of the meteorological variable. α is the turbulence modulation coefficient, which represents the amplification effect of turbulence intensity on the correlation of disturbance forces. Its value is obtained by fitting the Pearson correlation coefficient between turbulence intensity and disturbance forces in historical flight data, and its range is between 0 and 1.

[0069] The selection of kernel parameters is based on the following criteria: spatial length scale l s The time scale is determined based on the sensitivity of perturbation forces to changes in spatial position in historical flight trajectories. t The length scale of meteorological variables is determined based on the time interval when the time autocorrelation function of the disturbance decays to 1 / e. z The signal variance σ is calculated based on the variance ratio of the disturbance force changes under different meteorological conditions. f Based on the variance initialization of the disturbance force prediction sample set, the turbulence modulation coefficient α is obtained by fitting the correlation coefficient between turbulence intensity and disturbance force in historical data;

[0070] The Gaussian process potential force model is trained offline based on historical flight data, and online parameter updates are performed based on real-time environmental data during flight to complete spatiotemporal correlation learning.

[0071] The raw environmental data and real-time environmental data are input into the Gaussian process potential force model to generate disturbance force prediction results, which include mean estimation and uncertainty.

[0072] The perturbation prediction results are indexed with the starting state distribution, ending state distribution and reference stochastic process in the initial task dataset, and the perturbation prediction results are output.

[0073] The offline training process includes feature extraction from historical flight data, construction of input-output sample pairs, calculation of the kernel matrix, and hyperparameter optimization. Feature extraction involves time synchronization and normalization of spatial location, timestamp, flight altitude, airspeed, heading angle, wind speed and direction, turbulence intensity, temperature and humidity, and environmental obstacle information to form a spatiotemporal input vector. The output sample pairs are calculated by subtracting the predicted acceleration from the UAV flight dynamics model from the actual acceleration inferred from the position and velocity state estimates, and then subtracting the aerodynamic calculation results to obtain the estimated environmental disturbance force. Sample weighting uses weights generated from confidence intervals, assigning higher weights to samples with lower uncertainty to enhance the model's ability to fit high-confidence data. The kernel matrix calculation uses a spatiotemporal kernel function that introduces non-stationarity to reflect the differences between environmental obstacle information and terrain. The spatiotemporal kernel is composed of the product of a spatial correlation kernel and a temporal correlation kernel, and a modulation factor is set for turbulence intensity. Hyperparameter optimization uses marginal likelihood maximization and adds regularization constraints to prioritize the allocation of length scales in the sensitive dimensions of energy consumption cost and collision risk cost. The training process uses induced point sparsity approximation to reduce computational complexity.

[0074] In this embodiment, the generation of the optimal state distribution trajectory includes:

[0075] Based on the starting state distribution, ending state distribution, reference stochastic process, and disturbance prediction results, the Scholdinger bridge distribution evolution model is initialized. The starting state distribution and ending state distribution are used as boundary conditions, the reference stochastic process is used as the state evolution benchmark under no control input conditions, and the disturbance prediction results are introduced into the evolution process as external disturbance terms.

[0076] A multi-objective optimization model is constructed. The objective function of the multi-objective optimization model includes energy consumption cost and collision risk cost. The energy consumption cost is calculated by the cumulative thrust input and flight time, and the collision risk cost is calculated by the minimum distance between the optimal state distribution trajectory and the safety boundary of environmental obstacles. The energy consumption cost and collision risk cost are combined in a weighted manner to form the total objective function. The weight coefficients are dynamically adjusted according to the UAV mission type, remaining battery capacity and environmental congestion.

[0077] The Scholdinger bridge distribution evolution is iteratively solved based on a multi-objective optimization model. During the iteration process, the total objective function and the Scholdinger bridge constraint terms are minimized simultaneously to achieve a balance between smoothness, energy efficiency and safety of the state distribution trajectory. The optimal state distribution trajectory is output after convergence.

[0078] The calculation of energy consumption cost is based on the optimal state distribution trajectory. The trajectory is discretized into a control cycle sequence. In each cycle, the expected state change is converted into thrust input according to the UAV flight dynamics model. The scalar accumulation of the thrust input sequence is multiplied by the corresponding cycle duration to obtain the stage energy consumption. Then, the energy consumption cost is obtained by summing along the entire process and adding the total flight time as the time dimension weight. The calculation of collision risk cost is based on the safety boundary of environmental obstacles. The environmental obstacle information is converted into a set of boundaries with safety distances. The spatial position and uncertainty range of the optimal state distribution trajectory at each time step are evaluated by distance. The geometric minimum distance from the trajectory position to the nearest safety boundary is taken, and the distance is conservatively corrected by the distribution uncertainty range. When the corrected distance is negative, it is recorded as a crossing penalty. If it is positive, the risk score is accumulated in a segmented manner according to the inverse distance ratio or threshold. The collision risk cost is obtained by aggregation along the entire process.

[0079] The safety boundary is defined as a closed area formed by extending a safe distance outward from the physical outline of an obstacle. This safety distance is determined comprehensively based on factors such as the size of the UAV, positioning error, sensor measurement error, and response delay of the flight control system. It is used to ensure that there is still sufficient avoidance space between the UAV and the obstacle when there is deviation in the trajectory or external disturbance.

[0080] The method for calculating the minimum distance is based on the known geometric model of the obstacle safety boundary and the spatial position of the UAV at each time step on the trajectory. The distance from the UAV's position to each point on the safety boundary is calculated using the Euclidean distance formula, and the minimum value is taken as the minimum distance for that time step. If the safety boundary is a complex polygon or curved surface, the perpendicular distance from the UAV's position to each side of the polygon or each fragment of the curved surface can be calculated and compared with the endpoint distances. The minimum value is taken as the result of the minimum distance calculation.

[0081] In this embodiment, the adaptive adjustment of the Scholdinger bridge entropy regularization coefficient includes:

[0082] Uncertainty is extracted from the disturbance force prediction results and, together with the remaining battery capacity of the UAV, constitutes the adjustment input. Normalization and denoising are then performed to generate the adjustment input vector.

[0083] The Scholdinger bridge entropy regularization coefficient is updated based on the adjustment input vector within a preset upper and lower bound. The update step size and monotonic mapping rule are set to generate the current Scholdinger bridge entropy regularization coefficient.

[0084] The current Scholdinger bridge entropy regularization coefficient is applied to the Scholdinger bridge distribution evolution model to update the weights of the objective terms related to the reference stochastic process in the multi-objective optimization model. The weights of energy consumption cost and collision risk cost are synchronously scheduled according to the task configuration to generate the updated optimization problem.

[0085] During flight, a threshold triggering and hysteresis mechanism is executed according to the control cycle. When the uncertainty of the disturbance prediction result exceeds the threshold or the remaining battery capacity of the UAV is lower than the threshold, the Scholdinger bridge entropy regularization coefficient is increased. When the uncertainty and the remaining battery capacity of the UAV enter the hysteresis interval, the Scholdinger bridge entropy regularization coefficient is decreased.

[0086] The monotonic mapping rule is used to convert the uncertainty and remaining battery capacity in the control input vector into stable and predictable Scholdinger bridge entropy regularization coefficients. Specifically, the uncertainty and remaining battery capacity are first normalized to zero to one, and upper and lower bounds and a maximum step size for single-cycle changes are set. Then, using single-investigation table lookup or piecewise linear mapping, an increase in uncertainty is mapped to a non-decreasing change in coefficients, and a decrease in remaining battery capacity is mapped to a non-decreasing change in coefficients. When these two are combined, the more conservative approach is prioritized to generate the target coefficient. A dead zone and hysteresis interval are set between the target coefficient and the current coefficient; small fluctuations are not updated, and updates are only performed when thresholds are crossed, to suppress jitter. During updates, amplitude limiting and rate of change limiting are applied first, followed by upper and lower bound truncation to obtain the coefficients for the current control cycle. This rule ensures a monotonic relationship between input and output, while avoiding frequent adjustments through threshold, dead zone, and step size control, facilitating stable and consistent behavior between path smoothness and energy consumption tradeoffs.

[0087] In this embodiment, the process of mapping the optimal state distribution trajectory to the UAV control input sequence includes:

[0088] Receive discrete state points in the optimal state distribution trajectory, calculate the attitude adjustment amount and thrust requirement corresponding to each state point according to the UAV's flight dynamics model, and convert the calculation results into a set of control input commands;

[0089] According to the control cycle, the control input command set is sent to the UAV flight control system to execute the corresponding attitude adjustment and thrust control, so that the UAV flies according to the optimal state distribution trajectory.

[0090] During flight, the system receives real-time updates of the disturbance force prediction results from the Gaussian process potential force model. When the uncertainty of the disturbance force prediction results exceeds the threshold or the remaining battery capacity of the UAV is lower than the threshold, the Scholdinger bridge entropy regularization coefficient is increased. When the uncertainty and the remaining battery capacity of the UAV enter the hysteresis interval, the Scholdinger bridge entropy regularization coefficient is decreased, triggering the re-optimization process of the Scholdinger bridge distribution evolution model and recalculating the optimal state distribution trajectory.

[0091] The optimal state distribution trajectory obtained through re-optimization is mapped to a new set of control input instructions, which replaces the currently executed instructions within the control cycle, enabling a rapid response to sudden environmental changes.

[0092] The attitude adjustment is calculated based on the spatial position and attitude information of adjacent discrete state points in the optimal state distribution trajectory. Specifically, the three-dimensional position difference vector between the current state point and the target state point is first extracted. The required roll, pitch, and yaw angle changes are then calculated through coordinate transformation, and these angular changes are converted into corresponding attitude adjustment commands according to the UAV's attitude dynamics equations. The thrust requirement is calculated based on the UAV's flight dynamics model. The required velocity and acceleration at the state point are used as inputs, and the total thrust required is calculated using parameters such as mass, gravity, air resistance, and environmental disturbance forces. The thrust is then allocated to each propulsion unit. To ensure real-time performance and accuracy, the thrust allocation is also corrected using motor thrust curves and real-time environmental data, thereby generating precise thrust control commands.

[0093] Example 1:

[0094] To verify the feasibility of this invention in practice, it was applied to a multi-rotor unmanned aerial vehicle (UAV) performing a long-endurance autonomous flight mission in a complex environment. The mission area included undulating terrain, randomly distributed buildings and trees, and continuously changing weather conditions, including wind speed fluctuations, local turbulence, and changes in temperature and humidity. The UAV's mission was to take off from a designated starting point, avoid known and unknown obstacles, reach the destination while conserving energy as much as possible, and maintain a smooth and safe flight trajectory. Traditional path planning methods in such environments typically rely on static maps and preset paths, lacking real-time environmental modeling capabilities. This makes them prone to path deviations when weather changes or obstacle distribution shifts, leading to increased energy consumption and flight risks.

[0095] In application, this invention first collects state information such as the UAV's takeoff position, speed, and attitude, as well as environmental parameters such as wind speed and direction, turbulence intensity, temperature, and humidity in the mission area. Based on a multi-source sensor fusion algorithm, it generates starting and ending state distributions. Using a Gaussian process latent force model, it models and predicts real-time environmental disturbances, generating disturbance prediction results that include mean estimation and uncertainty. Combining the Schöringer bridge distribution evolution model and a multi-objective optimization model, it generates the optimal state distribution trajectory and maps it to attitude adjustment and thrust requirements, issuing control commands in real time. During flight, when the disturbance prediction uncertainty exceeds a threshold and the remaining battery capacity decreases, this invention triggers the Schöringer bridge optimization process, dynamically adjusting the entropy regularization coefficient to achieve real-time trajectory optimization, thereby effectively responding to sudden environmental changes.

[0096] To more intuitively demonstrate the performance differences, the results of this invention and the traditional path planning method are compared under the same test conditions. Table 1 is a statistical table of experimental data.

[0097] Table 1. Performance comparison of the present invention and traditional methods in complex environments.

[0098]

[0099] As shown in Table 1, the method of this invention outperforms traditional methods in several performance indicators under complex environments. Regarding trajectory accuracy, the average maximum trajectory deviation of this invention is only 0.42 meters, a reduction of 68.9% compared to traditional methods, significantly improving path tracking accuracy. In terms of energy consumption, the average energy consumption is reduced by 12.6%, with an even greater reduction of 18.5% under high disturbance conditions, demonstrating energy efficiency in complex weather environments. Regarding safety, the average minimum safety boundary distance is increased to 0.85 meters, a 73.5% improvement compared to traditional methods, effectively reducing the risk of collisions with obstacles. The mission completion rate is improved by 5.3%, and the average flight time is shortened by 10.1%, indicating that mission execution efficiency is improved while ensuring safety. The number of collision risk events is reduced by 75%, further demonstrating the stability and reliability of this invention in dynamic environments. These performance improvements are due to the invention's use of a Gaussian process latent force model combined with Scholdinger bridge optimization to predict and respond to changes in disturbance forces in real time, thus achieving an optimal balance between accuracy, energy efficiency, and safety.

[0100] Data analysis shows that this invention outperforms traditional methods in multiple indicators, including trajectory accuracy, energy consumption optimization, and safety assurance. The significant reduction in trajectory deviation is due to the invention's use of a Gaussian process latent force model to predict disturbance forces in real time and dynamically adjust the flight path within a Scholdinger bridge optimization framework, making the flight trajectory more closely match environmental changes. The core of the energy consumption reduction lies in the balance mechanism between smoothness and energy consumption in the multi-objective optimization model, enabling the UAV to complete tasks with lower energy consumption even under disturbed conditions. The improved safety boundary distance is attributed to real-time collision risk cost assessment and path correction mechanisms, effectively avoiding obstacles in highly disturbed environments. Furthermore, the improved mission completion rate and shortened flight time demonstrate the rapid response capability and stability of this invention in environmental perception and path replanning. In summary, this invention significantly outperforms traditional path planning methods in autonomous flight performance under complex dynamic environments, possessing advantages of high accuracy, low energy consumption, and high safety.

[0101] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for unmanned aerial vehicle (UAV) path planning, characterized in that, Includes the following steps: Perform environmental data acquisition and state modeling, and construct an initial task dataset that includes the starting state distribution, the ending state distribution, a reference stochastic process, raw environmental data, and real-time environmental data. Based on the Gaussian process potential force model, we model and infer the original environmental data and real-time environmental data, learn the spatiotemporal correlation of disturbance force using historical flight data and real-time environmental data, and generate disturbance force prediction results. Based on the starting state distribution, ending state distribution, reference stochastic process, and perturbation force prediction results, a Scholdinger bridge distribution evolution model is constructed to achieve a smooth transition from the starting state distribution to the ending state distribution while minimizing deviation from the reference stochastic process. At the same time, energy consumption cost and collision risk cost are introduced to form a multi-objective optimization model, and the optimal state distribution trajectory is output. Adaptive adjustment of the Scholdinger bridge entropy regularization coefficient is performed. The Scholdinger bridge entropy regularization coefficient adjusts the constraint strength of the distribution evolution process on the reference random process, and controls the smoothness of the state distribution trajectory and the energy consumption trade-off. The optimal state distribution trajectory is mapped to the control input sequence of the UAV. When the uncertainty of the disturbance prediction exceeds the set threshold, the Scholdinger bridge optimization is retried based on the real-time update result of the Gaussian process potential force model during flight.

2. The UAV path planning method according to claim 1, characterized in that, The construction of the initial task dataset includes: Collect multi-source sensor data from the UAV, synchronize and align the time stamps according to a unified time base, fuse the collected multi-source sensor data, use multi-sensor fusion algorithm to calculate the state estimates and confidence intervals of position and velocity, and generate the starting state distribution. Collect the endpoint state distribution and raw environmental data of the target area, and collect real-time environmental data during flight. Analyze and calculate the raw and real-time environmental data to generate environmental disturbance parameters. A reference stochastic process is established by combining the flight dynamics model of the UAV. By mapping environmental disturbance parameters to state transition probability distribution, the reference stochastic process is generated so that the state transition probability distribution can characterize the state change law of the UAV under the influence of the natural environment under the condition of no control input. The starting state distribution, ending state distribution, reference stochastic process, raw environmental data, and real-time environmental data are encapsulated into an initial task dataset.

3. The UAV path planning method according to claim 1, characterized in that, The generation of the disturbance prediction results includes: Raw environmental data and real-time environmental data are extracted from the initial task dataset to construct a disturbance prediction sample set; A Gaussian process potential force model was established, a spatiotemporal kernel function was selected, and kernel parameters and noise parameters were set. The spatiotemporal kernel function established the correlation between spatial location, time and meteorological variables, and the kernel parameters were initialized from historical flight data. The Gaussian process potential force model is trained offline based on historical flight data, and online parameter updates are performed based on real-time environmental data during flight to complete spatiotemporal correlation learning. The raw environmental data and real-time environmental data are input into the Gaussian process potential force model to generate disturbance force prediction results, which include mean estimation and uncertainty. The perturbation prediction results are indexed with the starting state distribution, ending state distribution and reference stochastic process in the initial task dataset, and the perturbation prediction results are output.

4. The UAV path planning method according to claim 1, characterized in that, The generation of the optimal state distribution trajectory includes: Based on the starting state distribution, ending state distribution, reference stochastic process, and disturbance prediction results, the Scholdinger bridge distribution evolution model is initialized. The starting state distribution and ending state distribution are used as boundary conditions, the reference stochastic process is used as the state evolution benchmark under no control input conditions, and the disturbance prediction results are introduced into the evolution process as external disturbance terms. A multi-objective optimization model is constructed. The objective function of the multi-objective optimization model includes energy consumption cost and collision risk cost. The energy consumption cost is calculated by the cumulative thrust input and flight time, and the collision risk cost is calculated by the minimum distance between the optimal state distribution trajectory and the safety boundary of environmental obstacles. The energy consumption cost and collision risk cost are combined in a weighted manner to form the total objective function. The weight coefficients are dynamically adjusted according to the UAV mission type, remaining battery capacity and environmental congestion. The Scholdinger bridge distribution evolution is iteratively solved based on a multi-objective optimization model. During the iteration process, the total objective function and the Scholdinger bridge constraint terms are minimized simultaneously to achieve a balance between smoothness, energy efficiency and safety of the state distribution trajectory. The optimal state distribution trajectory is output after convergence.

5. The UAV path planning method according to claim 1, characterized in that, The adaptive adjustment of the Scholdinger bridge entropy regularization coefficient includes: Uncertainty is extracted from the disturbance force prediction results and, together with the remaining battery capacity of the UAV, constitutes the adjustment input to generate the adjustment input vector. The Scholdinger bridge entropy regularization coefficient is updated based on the adjustment input vector within a preset upper and lower bound. The update step size and monotonic mapping rule are set to generate the current Scholdinger bridge entropy regularization coefficient. The current Scholdinger bridge entropy regularization coefficient is applied to the Scholdinger bridge distribution evolution model to update the weights of the objective terms related to the reference stochastic process in the multi-objective optimization model. The weights of energy consumption cost and collision risk cost are synchronously scheduled according to the task configuration to generate the updated optimization problem. During flight, threshold triggering and hysteresis mechanisms are executed according to the control cycle, only updating the Scholdinger bridge entropy regularization coefficient, without triggering path re-optimization.

6. The UAV path planning method according to claim 1, characterized in that, The process of mapping the optimal state distribution trajectory to the UAV control input sequence includes: Receive discrete state points in the optimal state distribution trajectory, calculate the attitude adjustment amount and thrust requirement corresponding to each state point according to the UAV's flight dynamics model, and convert the calculation results into a set of control input commands; According to the control cycle, the control input command set is sent to the UAV flight control system to execute the corresponding attitude adjustment and thrust control. During flight, the system receives real-time updates of the disturbance force prediction results from the Gaussian process potential force model. When the uncertainty of the disturbance force prediction results exceeds the threshold or the remaining battery capacity of the UAV is lower than the threshold, the Scholdinger bridge entropy regularization coefficient is increased. When the uncertainty and the remaining battery capacity of the UAV enter the hysteresis interval, the Scholdinger bridge entropy regularization coefficient is decreased, triggering the re-optimization process of the Scholdinger bridge distribution evolution model and recalculating the optimal state distribution trajectory. The optimal state distribution trajectory obtained through re-optimization is mapped to a new set of control input instructions, which are then used to replace the currently executed instructions within the control cycle.

Images