Unmanned aerial vehicle on-load imitated flight flight height planning method and system based on deep learning

Abstract

The invention discloses an unmanned aerial vehicle on-load terrain-imitated flight flight height planning method based on deep learning, and aims to solve the problem of model uncertainty-driven adaptive clearance margin and terrain-imitated flight height planning of verifiable hard constraints. A terrain high-confidence upper bound is obtained based on weighted quantile calibration of a forward speed, total time delay and an effective foresight distance, a self-adaptive clearance margin is generated by adopting mapping with monotonicity, and a hard security constraint of a flight height reference trajectory is constructed by combining security model predictive control containing control barrier function constraint with time delay compensation and online self-adaptive conservative. The technical effects that the clearance safety is improved under the complex terrain and the dynamic working condition, and trajectory tracking and energy consumption are considered are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of unmanned aerial vehicle (UAV) flight altitude planning, and in particular to a UAV load-carrying earth-following flight altitude planning method and system based on deep learning. BACKGROUND

[0002] In low-altitude earth-following flight tasks, the UAV needs to rely on forward-looking perception (cameras, laser radars, millimeter wave altimeters, etc.) and navigation fusion to plan the flight altitude in real time under complex terrain, load disturbance and wind field action, so as to ensure both obstacle clearance and trajectory tracking performance.

[0003] With the maturity of multi-source sensor time synchronization and external parameter registration technology, environment understanding and uncertainty estimation based on deep learning are gradually used for terrain height prediction; in the optimization and control aspects, model predictive control is widely used, and some research introduces probability constraints, risk metrics or control barrier functions to improve safety, and considers the time delay of the perception-to-execution link for state extrapolation and compensation.

[0004] However, for complex and rapidly changing working conditions, the existing technology still has the following shortcomings:

[0005] 1. Uncertainty calibration and safety upper bound lack verifiability: The confidence of most prediction models comes from experience or offline global calibration, which is difficult to conditionally adjust with forward speed, total time delay and effective forward-looking distance, and is prone to cause excessive conservatism of the clearance boundary or distortion in high-risk working conditions;

[0006] 2. Clearance margin design is rigid and difficult to adapt: Common methods use fixed or simplified margins, and do not systematically include monotonic mapping of factors such as terrain variance, local slope, speed and time delay, making it difficult to dynamically balance safety and performance;

[0007] 3. The planning level lacks verifiable hard constraints and time delay integrated processing in discrete time domain: Model predictive control based on probability or soft constraints cannot provide hard safety guarantees at each time, the integration of control barrier functions and flight altitude planning is insufficient, and the constraint satisfaction and feasibility maintenance mechanism is not perfect when the time delay exists.

[0008] Therefore, a flight altitude planning method and system that can solve the above-mentioned shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. SUMMARY

[0009] An object of the present application is to provide a deep learning-based unmanned aerial vehicle (UAV) load-carrying terrain-following flight altitude planning method and system. In view of the problems in the prior art that uncertainty prediction is difficult to conditionally calibrate, the clearance margin is fixed and rigid, the planning layer lacks verifiable hard constraints, and the time delay from perception to execution is not fully considered, a technical solution is provided for time synchronization and external parameter registration of multi-source observation, evidence-based deep learning terrain prediction, weighted quantile (conditionally conformal) calibration combining forward speed, total time delay, and effective forward distance to obtain a terrain height confidence upper bound, a self-adaptive clearance margin mapping with monotonicity, and a safety model predictive control with a control barrier function constraint and time delay compensation. Through online index-driven quantile level and mapping parameter closed-loop adaptive adjustment, the present application has the technical effects of realizing hard safety constraints at each time, improving clearance retention rate and obstacle crossing success rate, reducing violation event rate, and balancing energy consumption and trajectory tracking performance in complex terrain and dynamic working conditions.

[0010] According to an embodiment of the present application, a deep learning-based unmanned aerial vehicle (UAV) load-carrying terrain-following flight altitude planning method is provided, characterized by comprising:

[0011] S1, collect multi-source sensor observations, perform time synchronization and spatial external parameter registration, unify to a reference coordinate system, and obtain a synchronized and registered data sequence;

[0012] S2, input the data sequence, calculate the forward speed, total time delay from perception to execution, and effective forward distance, and arrange them at multiple discrete time points within a forward time window to form a conditional variable sequence;

[0013] S3, input the data sequence, use an evidence-based deep learning model to predict terrain height at discrete time points within the forward time window, obtain a terrain height mean sequence, a variance sequence, and a local slope absolute value sequence calculated from the mean sequence, and form a prediction set;

[0014] S4, input the prediction set, perform weighted quantile calibration based on the forward speed, total time delay, and effective forward distance on the residual between the historical true terrain height and the predicted mean within a sliding time window, obtain a terrain height high-confidence quantile upper bound at each discrete time point, and form a sequence. Combine the sequence with the terrain height mean sequence, variance sequence, local slope absolute value sequence, and forward speed, total time delay, and effective forward distance to form a calibration set;

[0015] S5, input the calibration set, use a mapping function to calculate the adaptive clearance margin at each discrete time point and form an adaptive clearance margin sequence, combine the sequence with the terrain height high-confidence quantile upper bound sequence to form a clearance set;

[0016] S6, inputting the clearance set, constructing and solving an optimization problem containing load-carrying flight dynamics constraints and hard safety inequalities to obtain a height reference trajectory;

[0017] S7, inputting the height reference trajectory, issuing and executing height or vertical speed instructions, counting online indicators, obtaining quantile levels and adjustment amounts of mapping function parameters according to threshold values, and taking the adjustment amounts as inputs of S4 and S5 in the next cycle to realize adaptive conservatism.

[0018] Optionally, step S1 is specifically:

[0019] Collecting multi-source sensor observations including at least forward-looking camera image data, laser radar point cloud data, millimeter wave radar height measurement data, inertial measurement unit data, global navigation satellite system positioning data and barometer height data;

[0020] Time synchronizing the observations to align each sensor data to a unified time reference, wherein the time synchronization adopts hardware trigger synchronization or software timestamp alignment, and in the case of inconsistent sampling rates or existence of transmission delay or processing delay, the alignment is realized through resampling, interpolation or buffer queue;

[0021] Spatial extrinsic parameter registration is performed on the observations to determine rigid pose transformation parameters of each sensor relative to a unified reference coordinate system, wherein the pose transformation parameters include rotation and translation, the unified reference coordinate system is a body coordinate system, an inertial coordinate system or a navigation coordinate system, and the pose transformation parameters are obtained through offline calibration or online estimation;

[0022] Accordingly, each observation is mapped to the unified reference coordinate system and arranged in time sequence to obtain a synchronized and registered data sequence as the input of subsequent steps.

[0023] Optionally, step S2 is specifically:

[0024] The time-synchronized and spatially extrinsically registered data sequence is inputted to calculate the forward speed of the unmanned aerial vehicle along the heading direction, the total time delay perceived and executed, and the effective forward-looking distance;

[0025] The forward speed is calculated by estimating the component of the speed of the unmanned aerial vehicle in the heading direction, wherein the heading direction is determined according to the solution of the attitude and the heading, and the speed estimation is obtained from the navigation fusion result of the data sequence;

[0026] The total time delay is calculated by counting the timestamps of the perception processing, fusion solution, prediction and planning, control command issuing and execution, etc., to obtain the time difference between the generation of the corresponding observation and the start of the execution of the control command, and to update the time difference over time;

[0027] The effective forward-looking distance is defined as the maximum forward-looking distance that can be reliably imaged or reliably ranged in the heading direction, and is determined by at least the field of view angle of the forward-looking sensor and the current flight height, and is calculated according to a geometric relationship under the condition that it does not exceed the nominal detection range of the forward-looking sensor;

[0028] The forward velocity, the total time delay, and the effective forward-looking distance at a plurality of discrete time instants within the forward time window are arranged to form a sequence of condition variables.

[0029] Optionally, step S3 is specifically:

[0030] The terrain height is predicted at a plurality of discrete time instants within the forward time window using an evidence-based deep learning model with the data sequence as input, and the mean terrain height and the terrain height variance at each discrete time instant are output.

[0031] The mean terrain height at each discrete time instant is sequentially arranged to form a sequence of mean terrain heights, and the terrain height variance at each discrete time instant is sequentially arranged to form a sequence of terrain height variances, and the absolute value of the local slope is calculated by adjacent difference or local fitting according to the sequence of mean terrain heights.

[0032] The sequence of mean terrain heights, the sequence of terrain height variances, and the sequence of absolute values of local slopes are combined to form a prediction set.

[0033] Optionally, step S4 is specifically:

[0034] A calibration sample set is constructed within a set sliding time window with the prediction set as input, the calibration sample set is composed of residuals of true terrain heights at historical time instants and corresponding mean terrain heights, the residual is defined as the true terrain height minus the corresponding mean terrain height, and the true terrain height is obtained by fusion calculation of a millimeter wave radar altimeter, a global navigation satellite system, and a barometer.

[0035] For each discrete time instant of the forward time window, the forward velocity, the total time delay, and the effective forward-looking distance in the sequence of condition variables are used to assign weights to the calibration sample, and the weights are non-negative values.

[0036] The residuals are quantile-estimated according to the weights to obtain a quantile calibration quantity at the discrete time instant, and the quantile calibration quantity is added to the mean terrain height at the discrete time instant to obtain a terrain height upper confidence quantile upper bound at the discrete time instant.

[0037] The terrain height upper confidence quantile upper bounds at each discrete time instant are sequentially arranged to form a sequence of terrain height upper confidence quantile upper bounds, and the sequence of terrain height upper confidence quantile upper bounds, the sequence of mean terrain heights, the sequence of terrain height variances, the sequence of absolute values of local slopes, and the forward velocity, the total time delay, and the effective forward-looking distance are collectively used to form a calibration set.

[0038] Optionally, step S5 specifically includes:

[0039] Using the calibration set as input, for each discrete moment within the forward time window, an adaptive clearance margin is calculated using a mapping function based on the terrain height variance, local slope absolute value, forward velocity, total delay, and effective forward look distance at that moment. The mapping function is monotonically non-decreasing with respect to terrain height variance, local slope absolute value, forward velocity, and total delay, and monotonically non-increasing with respect to effective forward look distance, and its parameters are adjustable.

[0040] Set lower and upper limits for the adaptive clearance margin, so that its value is between the lower and upper limits;

[0041] The adaptive net margins at each discrete time point are sequentially arranged into an adaptive net margin sequence, which is then combined with the upper bound sequence of the high confidence quantile of terrain height to form a net set.

[0042] Optionally, step S6 specifically includes:

[0043] Using the aforementioned airspace set as input, an optimization problem is constructed within a forward time window. The cost function of the optimization problem includes at least an energy consumption term and a reference track deviation term. The constraints include at least load-bearing flight dynamics model constraint parameters and a hard safety inequality. The load-bearing flight dynamics model constraint parameters include maximum climb rate, maximum descent rate, upper limit of attitude angle, lower limit of thrust margin, and upper limit of acceleration change rate. The hard safety inequality states that the flight altitude reference trajectory at each discrete time is not less than the sum of the upper limit of the high confidence quantile of terrain altitude at that time and the adaptive airspace margin.

[0044] Solve the optimization problem to generate an altitude reference trajectory that satisfies the constraints;

[0045] To compensate for the total time delay, the UAV state is extrapolated according to the time offset corresponding to the total time delay based on the loaded flight dynamics model to obtain the predicted state. The upper bound of the high confidence quantile of the terrain height at the time corresponding to the predicted state is then added to the optimization problem as the terrain height constraint after time delay compensation.

[0046] Construct a look-ahead barrier function such that the value of the barrier function at each discrete time step is equal to the flight altitude reference trajectory minus the upper bound of the corresponding terrain height high confidence quantile minus the corresponding adaptive airspace margin, and require that the weighted sum of the derivative of the barrier function and the gain function is not less than zero. Incorporate this barrier constraint into the optimization problem to ensure that the hard safety inequality is satisfied in time, and obtain the flight altitude reference trajectory.

[0047] Optionally, step S7 specifically includes:

[0048] Using the flight altitude reference trajectory as input, send altitude or vertical speed commands to the flight control interface and execute them;

[0049] During execution, online indicators are calculated based on the flight altitude reference trajectory and real-time observations, including minimum airspace hold rate, obstacle clearance success rate and violation rate, and the online indicators are statistically obtained within the sliding time window or the current execution cycle.

[0050] The online indicators are compared with preset thresholds, and the adjustment amount of the quantile level and the adjustment amount of the mapping function parameters are determined based on the comparison results. The adjustment amount is a numerical value and is limited to the preset upper and lower limits to ensure stability.

[0051] The adjustment amount of the quantile level is used to update the quantile level setting of the next cycle, and the adjustment amount of the mapping function parameter is used to update the mapping function parameter setting of the next cycle, so as to improve the conservatism under dangerous conditions and maintain performance under safe conditions.

[0052] In the next cycle, the updated quantile level and mapping function parameters are used as part of the inputs for steps S4 and S5, respectively, to achieve adaptive conservative closed-loop adjustment.

[0053] A deep learning-based UAV altitude planning system for terrain-following flight includes:

[0054] The perception alignment module is used to acquire multi-source observations, synchronize time and register extrinsic parameters, and output data sequences;

[0055] The condition variable module is used to calculate the forward velocity, total delay, and effective forward look distance of the data sequence, and output the condition variable sequence.

[0056] The prediction module is used to output a prediction set using evidence-based deep learning;

[0057] The calibration module is used to weight the residuals based on the condition variables within a sliding window, output the upper bound sequence of the high confidence quantile of terrain height, and form a calibration set with the prediction set and the condition variable sequence.

[0058] The margin module is used to map and generate an adaptive net margin sequence, which is combined with the upper bound sequence to form a net set;

[0059] The safety planning module is used to solve the flight altitude reference trajectory based on the aforementioned clearance set and load dynamic constraints, and to perform time delay compensation and look-ahead barrier constraints.

[0060] The adaptive module is used to issue height or vertical speed commands, collect online metrics, and output the adjustment amounts of the quantile level and mapping parameters for the next loop.

[0061] The beneficial effects of this invention are:

[0062] 1. Provide verifiable time-by-time hard safety assurance: Obtain a high confidence upper bound for terrain by using evidence-based deep learning prediction and weighted quantile calibration based on forward velocity, total delay, and effective forward look distance. Combined with monotonic adaptive clearance margin, control barrier function constraints, and delay compensation, apply a hard safety inequality to the flight altitude reference trajectory at the planning layer, significantly improving the minimum clearance hold-up rate and obstacle clearance success rate, and reducing the violation rate.

[0063] 2. Achieving an adaptive trade-off between safety and performance: The clearance margin mapping remains monotonically unchanged for terrain variance, local slope, forward speed, and total delay, and monotonically unchanged for effective forward sight distance. It also adjusts the quantile level and mapping parameters through online closed-loop adjustment of indicators, automatically increasing conservatism in hazardous conditions and reducing redundancy margin in safe conditions, thus balancing energy consumption and trajectory tracking accuracy.

[0064] 3. Improve coverage reliability and robustness under different working conditions: The time-sequential and conditional quantization calibration within the sliding time window effectively avoids the distortion of global unified calibration, enabling the terrain upper boundary to maintain stable coverage under conditions of speed changes, time delay fluctuations and limited line of sight, while also having stronger robustness to multi-source sensing alignment errors and load disturbances. Attached Figure Description

[0065] 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:

[0066] Figure 1 This is a flowchart of a deep learning-based method and system for planning the altitude of a drone carrying a terrain-following flight, proposed in this invention. Detailed Implementation

[0067] 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.

[0068] refer to Figure 1 A deep learning-based method for altitude planning of UAVs carrying terrain-following flight, characterized by comprising:

[0069] S1. Collect observations from multiple sensor sources, perform time synchronization and spatial extrinsic parameter registration, unify them to the reference coordinate system, and obtain the synchronized and registered data sequence;

[0070] S2. Using the data sequence as input, calculate the forward velocity, the total time delay from perception to execution, and the effective forward look-ahead distance, and arrange them according to multiple discrete moments within the forward time window to form a condition variable sequence.

[0071] S3. Using the data sequence as input, the evidence-based deep learning model is used to predict the terrain height at discrete moments in the forward time window, resulting in a terrain height mean sequence, variance sequence, and a local slope absolute value sequence calculated from the mean sequence, forming a prediction set.

[0072] S4. Using the predicted set as input, perform weighted quantile calibration on the residual between the historical true terrain height and the predicted mean within the sliding time window, based on forward velocity, total delay and effective forward sight distance, to obtain the upper bound of the high confidence quantile of terrain height at each discrete time and form a sequence. Combine this sequence with the terrain height mean sequence, variance sequence, local slope absolute value sequence and forward velocity, total delay and effective forward sight distance to form a calibration set.

[0073] S5. Using the calibration set as input, the adaptive airspace margin at each discrete time is calculated using a mapping function to form an adaptive airspace margin sequence. The sequence is then combined with the upper bound sequence of the high confidence quantile of terrain height to form an airspace set.

[0074] S6. Using the aforementioned clearance set as input, construct and solve an optimization problem that includes flight dynamics constraints and hard safety inequalities to obtain the flight altitude reference trajectory.

[0075] S7. Using the flight altitude reference trajectory as input, issue and execute altitude or vertical speed commands, statistically analyze online indicators, obtain the adjustment amount of quantile level and mapping function parameters based on the threshold, and use them as inputs to S4 and S5 in the next loop to achieve adaptive conservatism.

[0076] In this specific embodiment, S1 specifically refers to:

[0077] The system acquires multi-source observations from a forward-looking camera, lidar point clouds, millimeter-wave radar altimeter, inertial measurement unit, global navigation satellite system, and barometer. A unified reference time axis and a unified reference coordinate system are established as alignment benchmarks. Time synchronization and spatial extrinsic parameter registration are performed separately to form a data sequence for subsequent calculations. Time synchronization employs affine correction of the clocks of each sensor, mapping the original timestamps to the reference time axis. The correction model is written as follows:

[0078] ;

[0079] in For sensor index and Indicates camera, Indicates lidar, Indicates millimeter-wave radar altimeter, Indicating inertial measurement unit, Indicates Global Navigation Satellite System, Indicates barometer, For sensors The original timestamp, For sensors timestamp mapped to the reference timeline For sensors Clock drift coefficient, For sensors Time offset;

[0080] To unify the sampling rhythm, an equally spaced sampling grid is established on the reference time axis to drive resampling and interpolation processing. The sampling grid is written as:

[0081] ;

[0082] in For the first One reference sampling time, For the start time of the reference timeline, For reference sampling interval, The sample number is the index and its value ranges from 0 to... This refers to the total number of reference sampling points;

[0083] Spatial extrinsic parameter registration to unify the reference coordinate system For each sensor coordinate system, the target is... Estimate the rigid body pose and perform consistency verification. The mapping of 3D points from the sensor coordinate system to the reference coordinate system is written as:

[0084] ;

[0085] in In the sensor coordinate system 3D point vectors in To map to the reference coordinate system 3D point vectors From arrive of Rotation matrix From arrive of Translation vector;

[0086] In practice, hardware triggering is used to constrain initial alignment, while software timestamps and buffer queues are used to compensate for sampling rate inconsistencies and link latency, employing linear interpolation or hold-and-hold strategies. For filling in missing measurements, extrinsic parameters can be obtained through offline calibration and fine-tuned through small-range recursive estimation during the online phase. Finally, the observations from each path, synchronized with time and registered with spatial extrinsic parameters, are combined according to... The data is sequentially integrated into a unified data sequence to serve as input for subsequent steps.

[0087] In this specific embodiment, S2 specifically refers to:

[0088] Taking an aligned data sequence as input, discrete time points are set at equal intervals within a given forward time window, and the required condition variables are calculated time-by-time. The discrete time points are written as:

[0089] ;

[0090] in For the first discrete time, For the start time of the time window, For time step, For time indexing, End index;

[0091] The forward velocity is given by the projection of the velocity obtained from navigation fusion onto the forward direction, written as:

[0092] ;

[0093] in For the first The forward velocity at that moment, For the first The heading unit vector at each moment, Reference coordinate system Next The velocity vector at each moment Indicates transpose, This refers to the unified reference coordinate system established in step S1;

[0094] The total perceived execution latency is calculated as the difference between the command execution time and the observation generation time, written as:

[0095] ;

[0096] in For the first The total delay at each moment The timestamp at which the control command begins execution at that moment. The timestamp generated for observations related to that moment;

[0097] The smaller of the effective forward-looking distance, nominal detection range, and geometric visibility is taken as:

[0098] ;

[0099] in For the first Effective forward sight distance at any given moment The nominal detection range of the forward-looking sensor, The first is calculated from the relationship between flight altitude, sensor attitude, and field of view geometry. Geometric forward distance at each moment It is a binary minimum value operator;

[0100] Finally, the above quantities are arranged in chronological order to form a sequence of condition variables:

[0101] and ;

[0102] in For condition variable sequences, For the first The condition variable triples at each time point are used to drive subsequent uncertainty calibration and flight altitude planning calculations.

[0103] In this specific embodiment, S3 specifically refers to:

[0104] Using the aligned data sequence as input, the uncertainty of terrain elevation is predicted for each discrete time point within the forward time window, generating a prediction set for subsequent calibration and planning. Specifically, for the ... discrete time ;

[0105] in For time index and For the first time window At that moment, End index;

[0106] Constructing feature blocks from multi-source observations ;

[0107] in To integrate cameras, lidar, millimeter-wave radar altimeters, inertial measurement units, global navigation satellite systems, and barometers, etc. The feature vector or tensor obtained by fusing observations at a location;

[0108] And with evidence-based deep learning models Perform parametric prediction of terrain height, and write the model output as follows:

[0109] ;

[0110] in In order to be in The predicted average terrain height at the location In order to be in Variance of terrain height prediction at location For parameters Evidence-based deep learning models It is a set of trainable parameters;

[0111] The predicted mean and predicted variance at each time point are arranged into a sequence:

[0112] and ;

[0113] The absolute value of the local slope is calculated based on the change in the adjacent predicted mean to characterize the sensitivity of the terrain to undulation. The adjacent difference form is written as... ;

[0114] in In order to be in The absolute value of the local slope at the location, for and The discrete spacing along the flight direction (a positive number determined by both platform motion and sampling step size) is used to combine the three types of quantities into a prediction set:

[0115] ;

[0116] in This is the input set used for subsequent weighted quantile calibration and net-space mapping.

[0117] In this specific embodiment, S4 specifically refers to:

[0118] Using the prediction set and the condition variable sequence as input, conditional weighted quantile calibration of historical residuals is performed around each forward discrete time step within a set sliding time window to obtain a high-confidence upper bound for terrain height and form a calibration set for subsequent mapping and planning; specifically, for the ... At that moment ;

[0119] in For time indexing, For the first time window discrete time, End index and ;

[0120] Building a sliding window index set Used to collect recent historical samples, for each Calculate the residuals:

[0121] ;

[0122] in For the first The residuals of a historical sample For the first The actual terrain elevation at any given time (obtained by fusing millimeter-wave radar altimeter, global navigation satellite system, and barometer data). For the first The average predicted terrain height at each time point;

[0123] To achieve conditional weighting based on forward velocity, total delay, and effective forward look distance, weights are defined based on similarity:

[0124] ;

[0125] in For the first The sample relative to the first Non-negative weights at each moment Forward velocity difference, For total delay difference, For effective forward sight distance difference, Forward speed, To perceive the total execution latency, For effective forward sight distance, The weighting scale hyperparameter is used to adjust the sensitivity to differences;

[0126] Define quantile calibration values ​​on the weighted samples:

[0127] ;

[0128] in For the first Quantile calibration value at each moment For the infimum operator, For the real number field, For indicator functions, Quantitative level;

[0129] The upper bound of the high-confidence quantile for terrain height is obtained by superimposing the quantile calibration value with the predicted mean:

[0130] ;

[0131] in For the first The upper bound of the high confidence quantile for terrain elevation at each moment, For the first The average predicted terrain height at each time point;

[0132] Ultimately and as well as according to Combined to form a calibration set:

[0133] Adaptive headroom mapping and safety planning to drive the next step.

[0134] In this specific embodiment, S5 specifically includes:

[0135] Using the calibration set as input, the adaptive airspace margin is calculated time-by-time within the forward time window based on terrain uncertainty and working conditions, and combined with the high confidence upper bound of terrain height to form the airspace set.

[0136] Specifically, for the first discrete time ;

[0137] in For the first time window discrete time, For time indexing;

[0138] Select the terrain height variance at this moment absolute value of local slope Forward speed Perceive the total execution latency and effective forward sight distance As input to the mapping, a monotonic, untruncated margin mapping is constructed, and saturation is performed using upper and lower bounds to obtain the final margin. The untruncated form is written as:

[0139] ;

[0140] in For the first Uncut margin at each moment, Non-negative adjustable weight parameters are used to control the contribution of each input to the margin. For the square root operator, Variance for predicting terrain height For the absolute value of the local slope, Forward speed, For total latency, For effective forward sight distance, To prevent small positive quantities with a denominator of zero;

[0141] when When the structure is relative to and Monotonous and undiminished The monotonicity does not increase with increasing volume, thus satisfying the monotonicity constraint in the claims;

[0142] The uncrunted result is then saturated to obtain the final adaptive net clearance margin:

[0143] ;

[0144] in For the first Adaptive headroom at any given moment and These are the lower and upper margin constants, respectively. For saturation operators, and These are the binary minimum and binary maximum operators, respectively;

[0145] In chronological order Composition of adaptive net clearance sequence and the upper bound sequence of high confidence quantiles for terrain height. Combined to form a set of net spaces To facilitate subsequent hard safety constraints, a safe terrain baseline can also be defined. and its sequence ;

[0146] in For the first The upper bound of the high confidence quantile for terrain elevation at each moment, The weighting parameter is the sum of the upper bound and the margin, used to constrain the flight altitude to not be lower than the baseline. and upper and lower limits It is adjustable within a preset range to allow for an online trade-off between safety and performance.

[0147] In this specific embodiment, S6 specifically refers to:

[0148] Using the clearance set as the constraint input, a discrete time-domain optimization is established within the forward time window, while simultaneously considering energy consumption and reference tracking, longitudinal dynamics with load, envelope constraints, hard safety inequalities after time delay compensation, and look-ahead barrier function constraints to generate the flight altitude reference trajectory.

[0149] First, we plan on discrete time points, and write the time scale as follows: ;

[0150] in For the first discrete time, For the start time of the time window, For discrete time step, For time indexing, End index;

[0151] The decision variable is selected as And using quadratic cost to measure tracking and energy consumption, the objective function is written as follows:

[0152] ;

[0153] in To optimize costs, For time indexing Summation operator To track weights, For energy consumption weight, For the first At that moment, the flight altitude reference For the first Expected altitude at that moment For the first Vertical acceleration at each moment;

[0154] The longitudinal discrete dynamics adopts a minimum-order linear model:

[0155] ;

[0156] in For the first At that moment, the flight altitude reference For the first Vertical velocity at each moment For the first Vertical velocity at each moment;

[0157] The load envelope is primarily constrained by rate and jerk constraints.

[0158] ;

[0159] in For maximum climb rate, For the maximum rate of decline, For absolute value operators, The upper limit of vertical acceleration, attitude angle and thrust margin constraints are also incorporated into the inequality, but for the sake of simplicity, they are expressed in words here.

[0160] The hard safety inequality for headroom uses the sum of the high-confidence terrain upper bound and the adaptive headroom margin as a baseline and considers the total link delay for timing extrapolation. First, the discrete delay step number and compensation index are defined:

[0161] ;

[0162] in For the first The number of discrete time delay steps at each time point, For rounding operators, For the first The total latency from perception to execution at any given moment For the constraint time index after delay compensation, It is a binary minimum value operator;

[0163] Based on this, write out the hard safety constraints. ;

[0164] in For compensation index The corresponding terrain height high confidence quantile upper bound, For the first Adaptive headroom margin at any given moment;

[0165] To enhance the verifiability of time series, a look-ahead barrier function is introduced:

[0166] It also requires that its discrete "derivative plus gain" be non-negative to suppress security boundary intrusion, written as ;

[0167] in For the first The barrier function value at each time point, Forward gain coefficient, For the first Upper bound of the high confidence quantile for terrain elevation at each moment;

[0168] In summary, this constitutes "minimization" Furthermore, it satisfies the hard safety inequalities after dynamics, envelope, and time delay compensation, as well as the look-ahead barrier constraint, using a convex quadratic programming problem or a nonlinear quadratic programming problem with linear constraints, and is solved by a numerical optimizer in each programming cycle. As a reference trajectory for flight altitude.

[0169] In this specific embodiment, S7 specifically refers to:

[0170] Using the flight altitude reference trajectory as the execution input, the flight control interface selects to issue altitude or vertical speed commands. During real-time execution, a sliding window is used to statistically analyze online safety and performance indicators, and based on this, the amplitude limit update of the quantile level and mapping parameters is generated for the setting of steps S4 and S5 in the next cycle.

[0171] To standardize the measurement of airspace boundaries, the airspace margin is defined as:

[0172] ;

[0173] in For the first The clearance margin at any given moment For the first At that moment, the flight altitude reference For the first The upper bound of the high confidence quantile for terrain elevation at each moment, For the first Adaptive headroom margin at any given moment;

[0174] In length sliding window The minimum clearance rate, obstacle clearance success rate, and violation rate are statistically analyzed and written as follows:

[0175] and and ;

[0176] in For minimum headroom retention rate, To improve the success rate of obstacle crossing, For the rate of violations, For the number of samples in the window, For the set of time indexes within the window, For sets Internal index Summation operator For indicator functions, The number of times an obstacle can be overcome within the window while maintaining clearance. This represents the total number of obstacle crossing events within the window.

[0177] The above indicators are compared with the threshold to form a non-negative risk indicator:

[0178] ;

[0179] in For risk indicators, For weight, For preset threshold, It is a positive part operator and It is a binary maximum value operator;

[0180] Quantile levels and mapping parameters are updated using amplitude-limited additive updates to ensure robustness. The update law is as follows:

[0181] and ;

[0182] in For the next cycle, the quantile level used in step S4, For the current quantile level, For quantile gain, For quantile upper and lower limits, For the mapping parameter vector used in step S5 of the next loop, For the current mapping parameter vector, For parameter gain, To and Same-dimensional all-one vectors For the parameter, element-wise upper and lower bound vectors, For amplitude limiting operator and It is a binary minimum value operator;

[0183] Through the above closed loop or Below the threshold or When the value exceeds a threshold, the conservatism of the quantile level and mapping parameters is automatically increased, while maintaining parameter stability under safe operating conditions. and These will be used as inputs for the next loop.

[0184] 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.

[0185] This application addresses the technical challenge of simultaneously satisfying safety and performance requirements at the airspace boundary in low-altitude terrain-following flight due to prediction uncertainties, link delays, and limited forward look-ahead distances. It constructs a time-by-time verifiable hard safety constraint by employing a combination of calibration, margin, planning, and evaluation. Specifically, starting with multi-source perception alignment, evidence-based deep learning outputs the mean and variance of terrain height and calculates the local slope. Within a sliding time window, weighted quantile calibration is performed based on forward velocity, total delay, and effective forward look-ahead distance to obtain the upper bound of the high-confidence quantile for terrain height. An adaptive airspace margin is then generated using a monotonic mapping, and delay compensation is incorporated into the safety model predictive control with control barrier function constraints. The "upper bound + margin" is used as a time-by-time hard inequality to plan the flight altitude reference trajectory. This combination improves the minimum airspace hold rate and obstacle clearance success rate, reduces the violation rate, and balances energy consumption and trajectory tracking. During the execution phase, closed-loop adjustment based on online indicators and thresholds automatically increases conservatism in hazardous conditions and reduces redundancy in safe conditions, demonstrating stable and verifiable technical effects.

[0186] In terms of algorithm structure, this application makes targeted improvements to the core links to better address the above-mentioned technical problems and enhance the technical effect: (1) The sliding window calibration with conditional weighted quantiles is used to replace the offline global calibration, so that the terrain upper bound is dynamically updated with speed, time delay and forward look distance and the coverage reliability is maintained, avoiding global calibration distortion; (2) A margin mapping with uniform variance, slope, forward speed and total time delay, uniform effective forward look distance and upper and lower limit saturation is designed to avoid parameter adjustment direction errors and improve the adaptive trade-off between engineering controllability and performance-safety; (3) The time delay compensation and control barrier function are incorporated into the discrete time domain constraints of MPC to ensure the timing satisfaction and feasibility maintenance of hard safety inequalities. The limit update of quantile level and mapping parameters by the execution adaptive module forms a closed-loop robust mechanism, which further consolidates the safety of the airspace and improves the overall planning performance under complex terrain and dynamic working conditions.

Claims

1. A deep learning-based method for altitude planning during unmanned aerial vehicle (UAV) payload-based terrain-following flight, characterized in that, include: S1. Collect observations from multiple sensor sources, perform time synchronization and spatial extrinsic parameter registration, unify them to the reference coordinate system, and obtain the synchronized and registered data sequence; S2. Using the data sequence as input, calculate the forward velocity, the total time delay from perception to execution, and the effective forward look-ahead distance, and arrange them according to multiple discrete moments within the forward time window to form a condition variable sequence. S3. Using the data sequence as input, the evidence-based deep learning model is used to predict the terrain height at discrete moments in the forward time window, resulting in a terrain height mean sequence, variance sequence, and a local slope absolute value sequence calculated from the mean sequence, forming a prediction set. S4. Using the predicted set as input, perform weighted quantile calibration on the residual between the historical true terrain height and the predicted mean within the sliding time window, based on forward velocity, total delay and effective forward sight distance, to obtain the upper bound of the high confidence quantile of terrain height at each discrete time and form a sequence. Combine this sequence with the terrain height mean sequence, variance sequence, local slope absolute value sequence and forward velocity, total delay and effective forward sight distance to form a calibration set. S5. Using the calibration set as input, the adaptive airspace margin at each discrete time is calculated using a mapping function to form an adaptive airspace margin sequence. The sequence is then combined with the upper bound sequence of the high confidence quantile of terrain height to form an airspace set. S6. Using the aforementioned clearance set as input, construct and solve an optimization problem that includes flight dynamics constraints and hard safety inequalities to obtain the flight altitude reference trajectory. S7. Using the flight altitude reference trajectory as input, issue and execute altitude or vertical speed commands, statistically analyze online indicators, obtain the adjustment amount of quantile level and mapping function parameters based on the threshold, and use them as inputs to S4 and S5 in the next loop to achieve adaptive conservatism.

2. The method for altitude planning of UAVs with payload-based terrain-following flight based on deep learning according to claim 1, characterized in that, S1 specifically refers to: The data collection includes multi-source sensor observations such as forward-looking camera image data, lidar point cloud data, millimeter-wave radar altitude measurement data, inertial measurement unit data, global navigation satellite system positioning data, and barometer altitude data. The observations are time-synchronized to align the data from each sensor to a unified time reference. The time synchronization is achieved by hardware-triggered synchronization or software timestamp alignment. In the event of inconsistent sampling rates or transmission or processing delays, alignment is achieved through resampling, interpolation, or buffer queues. Spatial extrinsic parameter registration is performed on the observations to determine the rigid body pose transformation parameters of each sensor relative to a unified reference coordinate system. The pose transformation parameters include rotation and translation. The unified reference coordinate system is a body coordinate system, an inertial coordinate system, or a navigation coordinate system. The pose transformation parameters are obtained through offline calibration or online estimation. Accordingly, each observation is mapped to the unified reference coordinate system and arranged in chronological order to obtain a synchronized and registered data sequence as input for subsequent steps.

3. The method for altitude planning of UAVs with payload-based terrain-following flight based on deep learning according to claim 1, characterized in that, S2 specifically refers to: Using a data sequence that has been time-synchronized and spatially registered with extrinsic parameters as input, the forward velocity of the UAV along the heading, the total latency of sensing and execution, and the effective forward-looking distance are calculated. The forward velocity is obtained by estimating the component of the UAV's velocity in the heading direction, which is determined based on the attitude and heading calculation results. The velocity estimate is obtained from the navigation fusion results of the data sequence. The total delay is calculated by statistically analyzing the timestamps of the sensing processing, fusion calculation, prediction and planning, control command issuance and execution, etc., to obtain the time difference between the generation of the corresponding observation and the start of the execution of the control command, and is updated over time. The effective forward-looking distance is defined as the maximum forward distance that can reliably image or reliably measure distance in the heading direction. It is determined by the field of view of the forward-looking sensor and the current flight altitude, and is calculated based on geometric relationships under the condition that it does not exceed the nominal detection range of the forward-looking sensor. The forward velocity, the total delay, and the effective forward look-ahead distance are arranged at multiple discrete moments within the forward time window to form a sequence of conditional variables.

4. The method for altitude planning of UAVs with payload-based terrain-following flight based on deep learning according to claim 1, characterized in that, S3 specifically refers to: Using the data sequence as input, an evidence-based deep learning model is used to predict terrain height at multiple discrete moments within a forward time window, and the mean and variance of terrain height at each discrete moment are output. The mean elevation values ​​at each discrete time point are arranged into a mean elevation value sequence, and the variance of elevation at each discrete time point is arranged into a variance elevation value sequence. The absolute value of the local slope is calculated based on the mean elevation value sequence through adjacent difference or local fitting. The mean sequence of terrain height, the variance sequence of terrain height, and the absolute value sequence of local slope are combined to form a prediction set.

5. The method for altitude planning of UAVs with payload-based terrain-following flight according to claim 1, characterized in that, S4 specifically refers to: Using the predicted set as input, a calibration sample set is constructed within a set sliding time window. The calibration sample set consists of the residual between the actual terrain height at historical moments and the corresponding mean terrain height. The residual is defined as the actual terrain height minus the corresponding mean terrain height. The actual terrain height is obtained by the fusion calculation of a millimeter-wave radar altimeter, a global navigation satellite system, and a barometer. For each discrete moment of the forward time window, weights are assigned to the calibration samples based on the forward velocity, total delay, and effective forward look distance in the condition variable sequence, and the weights are non-negative values. The residuals are quantile statistics are performed according to the weights to obtain the quantile calibration value at the discrete time. The quantile calibration value is then added to the mean terrain height at the discrete time to obtain the upper bound of the high confidence quantile of the terrain height at the discrete time. The upper bounds of the high confidence quantiles of terrain height at each discrete time are arranged in order to form a sequence of upper bounds of the high confidence quantiles of terrain height. This sequence, together with the mean sequence of terrain height, the variance sequence of terrain height, the absolute value sequence of local slope, and the forward velocity, total delay, and effective forward sight distance, forms a calibration set.

6. The method for altitude planning of UAVs with payload-based terrain-following flight based on deep learning according to claim 1, characterized in that, S5 specifically refers to: Using the calibration set as input, for each discrete moment within the forward time window, an adaptive clearance margin is calculated using a mapping function based on the terrain height variance, local slope absolute value, forward velocity, total delay, and effective forward look distance at that moment. The mapping function is monotonically non-decreasing with respect to terrain height variance, local slope absolute value, forward velocity, and total delay, and monotonically non-increasing with respect to effective forward look distance, and its parameters are adjustable. Set lower and upper limits for the adaptive clearance margin, so that its value is between the lower and upper limits; The adaptive net margins at each discrete time point are sequentially arranged into an adaptive net margin sequence, which is then combined with the upper bound sequence of the high confidence quantile of terrain height to form a net set.

7. The method for altitude planning of UAV payload-based terrain-following flight based on deep learning according to claim 1, characterized in that, S6 specifically refers to: Using the aforementioned airspace set as input, an optimization problem is constructed within a forward time window. The cost function of the optimization problem includes at least an energy consumption term and a reference track deviation term. The constraints include at least load-bearing flight dynamics model constraint parameters and a hard safety inequality. The load-bearing flight dynamics model constraint parameters include maximum climb rate, maximum descent rate, upper limit of attitude angle, lower limit of thrust margin, and upper limit of acceleration change rate. The hard safety inequality states that the flight altitude reference trajectory at each discrete time is not less than the sum of the upper limit of the high confidence quantile of terrain altitude at that time and the adaptive airspace margin. Solve the optimization problem to generate an altitude reference trajectory that satisfies the constraints; To compensate for the total time delay, the UAV state is extrapolated according to the time offset corresponding to the total time delay based on the loaded flight dynamics model to obtain the predicted state. The upper bound of the high confidence quantile of the terrain height at the time corresponding to the predicted state is then added to the optimization problem as the terrain height constraint after time delay compensation. Construct a look-ahead barrier function such that the value of the barrier function at each discrete time step is equal to the flight altitude reference trajectory minus the upper bound of the corresponding terrain height high confidence quantile minus the corresponding adaptive airspace margin, and require that the weighted sum of the derivative of the barrier function and the gain function is not less than zero. Incorporate this barrier constraint into the optimization problem to ensure that the hard safety inequality is satisfied in time, and obtain the flight altitude reference trajectory.

8. The method for altitude planning of UAVs with payload-based terrain-following flight based on deep learning according to claim 1, characterized in that, S7 specifically refers to: Using the flight altitude reference trajectory as input, send altitude or vertical speed commands to the flight control interface and execute them; During execution, online indicators are calculated based on the flight altitude reference trajectory and real-time observations, including minimum airspace hold rate, obstacle clearance success rate and violation rate, and the online indicators are statistically obtained within the sliding time window or the current execution cycle. The online indicators are compared with preset thresholds, and the adjustment amount of the quantile level and the adjustment amount of the mapping function parameters are determined based on the comparison results. The adjustment amount is a numerical value and is limited to the preset upper and lower limits to ensure stability. The adjustment amount of the quantile level is used to update the quantile level setting of the next cycle, and the adjustment amount of the mapping function parameter is used to update the mapping function parameter setting of the next cycle, so as to improve the conservatism under dangerous conditions and maintain performance under safe conditions. In the next cycle, the updated quantile level and mapping function parameters are used as part of the inputs for steps S4 and S5, respectively, to achieve adaptive conservative closed-loop adjustment.

9. A deep learning-based UAV altitude planning system for payload-following terrain-following flight, used to execute the deep learning-based UAV altitude planning method for payload-following terrain-following flight as described in any one of claims 1 to 8, comprising: The perception alignment module is used to acquire multi-source observations, synchronize time and register extrinsic parameters, and output data sequences; The condition variable module is used to calculate the forward velocity, total delay, and effective forward look distance of the data sequence, and output the condition variable sequence. The prediction module is used to output a prediction set using evidence-based deep learning; The calibration module is used to weight the residuals based on the condition variables within a sliding window, output the upper bound sequence of the high confidence quantile of terrain height, and form a calibration set with the prediction set and the condition variable sequence. The margin module is used to map and generate an adaptive net margin sequence, which is combined with the upper bound sequence to form a net set; The safety planning module is used to solve the flight altitude reference trajectory based on the aforementioned clearance set and load dynamic constraints, and to perform time delay compensation and look-ahead barrier constraints. The adaptive module is used to issue height or vertical speed commands, collect online metrics, and output the adjustment amounts of the quantile level and mapping parameters for the next loop.

Images