Unmanned aerial vehicle moving trajectory intelligent evolution learning method based on virtual simulation

Abstract

The invention discloses a virtual simulation-based intelligent evolutionary learning method for a running trajectory of an unmanned aerial vehicle, and the method comprises the steps: reading historical trajectories from a trajectory library of the unmanned aerial vehicle, and carrying out the sorting according to a set evaluation standard; setting a threshold to distinguish high-quality samples and low-quality samples, adding the high-quality samples and part of the low-quality samples extracted according to a set rule into a sample set, and dividing the sample set into a training set and a test set according to a proportion; defining a loss function considering the comprehensive cost, and determining a comprehensive evaluation index through system simulation; and training the evolutionary learning model, controlling the aircraft in a simulation environment, and outputting a new aircraft trajectory sample. And if a system simulation evaluation index requirement is met, ending the process, otherwise, combining a new track generated by simulation with the previous track to form a new track library, and returning to carry out iterative optimization again. The method can effectively improve the efficiency and precision of aircraft trajectory optimization, and reduces the training cost at the same time.

Description

Technical Field

[0001] The present invention relates to an intelligent evolutionary learning method for the flight trajectory of an unmanned aerial vehicle based on virtual simulation, belonging to the technical field of aircraft control and machine learning. Background Art

[0002] With the rapid development of unmanned aerial vehicles and automatic control systems, the optimization of flight trajectories has become an important means to improve the performance of aircraft. In the prior art, the trajectory optimization of aircraft mainly relies on traditional control algorithms and single optimization strategies, and it is difficult to achieve efficient and accurate control in complex and changeable environments. As a machine learning technology, evolutionary learning can dynamically optimize trajectories and iteratively improve them in a simulation environment. However, how to effectively select samples, control training costs, and evaluate model performance remains a key issue to be solved. Summary of the Invention

[0003] Object of the Invention: To provide an intelligent evolutionary learning method for the flight trajectory of an unmanned aerial vehicle based on virtual simulation, aiming to achieve dynamic optimization of the aircraft trajectory through trajectory data sorting, high- and low-quality sample sampling strategies, loss function optimization, and simulation evaluation. The present invention can improve the efficiency and accuracy of trajectory control, effectively control training costs at the same time, and is applicable to various complex flight environments and tasks.

[0004] Technical Solution: To solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0005] An intelligent evolutionary learning method for the flight trajectory of an unmanned aerial vehicle based on virtual simulation, the method comprising:

[0006] (1) Establish an aircraft trajectory library including several trajectory segments, and divide all trajectory segments into two categories: high-quality and low-quality;

[0007] (2) Select trajectory segments from the trajectory library to obtain a trajectory segment sample set for training an evolutionary learning model;

[0008] (3) Train the evolutionary learning model based on the sample set selected in step (2);

[0009] (4) Put the trained evolutionary learning model into a simulation environment for simulation testing. If the aircraft satisfies the requirements of the preset simulation comprehensive evaluation index when performing a flight mission according to the predicted trajectory output by the trained evolutionary learning model, the algorithm ends, and the trained evolutionary learning model is solidified; otherwise, return to step (1), and supplement the predicted trajectory output by the trained evolutionary learning model into the trajectory library to form a new trajectory library.

[0010] Preferably, the calculation formula for the ratio P of high-quality trajectory segments to low-quality trajectory segments in step (1) is:

[0011]

[0012] Where: N total is the total number of trajectory segments; error(i) is the prediction error of the evolutionary learning model for the i-th trajectory segment in the trajectory; error max is the maximum prediction error among all trajectory segments in the trajectory; λ is the sensitivity parameter; ω cluster (i) is the weight value corresponding to whether the i-th trajectory segment in the trajectory belongs to a high-quality trajectory segment or a low-quality trajectory segment.

[0013] Preferably, the number of samples N in the sample set used to train the evolutionary learning model in step (2) sample is calculated by the formula:

[0014]

[0015] Where: T max is the maximum acceptable training time; Step is the average training time per trajectory segment sample; R exp is the expected relative improvement ratio of model performance.

[0016] Preferably, the method of selecting trajectory segments from the trajectory library in step (2) includes:

[0017] (2.1) Select all high-quality trajectory segments in the trajectory library into the sample set;

[0018] (2.2) Extract a certain number of low-quality trajectory segments in descending order of extraction probability to meet the sample number requirement of the sample set;

[0019] Where the calculation formula of the extraction probability includes:

[0020]

[0021] Where: P sample (j) is the extraction probability of the j-th low-quality trajectory segment; Q(j) is the Q value of the j-th low-quality sample; N low is the total number of low-quality trajectory segments.

[0022] Preferably, training the evolutionary learning model in step (3) further includes: dynamically dividing the sample set to obtain a training set and a test set, where the calculation formula of the ratio R of the number of trajectory segments in the training set and the test set is:

[0023]

[0024] Where: R init is the preset initial ratio of the training set and the test set; α is the adjustment parameter; N epochs is the number of iterations of the current training.

[0025] Preferably, the expression of the loss function for training the evolutionary learning model in step (3) is:

[0026] L = α1·L position + α2·L energy + α3·L time + α4·L smooth ++ α5·L collision

[0027] where: α1, α2, α3, α4, α5 are weight coefficients; L position is the trajectory position error loss, p n is the actual position of the nth trajectory point in the trajectory segment sample, is the predicted position of the nth trajectory point, and N is the number of trajectory points in the trajectory segment sample; L energy is the energy consumption loss, E n is the energy consumption of the nth trajectory point, and C energy is the cost per unit energy; L time is the time consumption loss, L time = T total ·C time , T total is the total time required for the aircraft to complete a flight mission, and C time is the cost per unit time; L smooth is the path smoothness loss, where, v n is the velocity vector of the nth trajectory point, v n+1 is the velocity vector of the (n - 1)th trajectory point; L collision is the collision risk loss, P collision (p n ) is the collision probability of the nth trajectory point with the obstacle.

[0028] Preferably, the calculation formula of the simulation comprehensive evaluation index in step (4) is as follows:

[0029] E = β1·E accuracy + β2·E safety + β3·E efficiency + β4·E stability + β5·E robustness

[0030] where: β1, β2, β3, β4, β5 are weight coefficients; E accuracy represents the accuracy index of the trajectory, M correctis the number of trajectory points with correct prediction, M total is the total number of trajectory points; E safety represents the flight safety index M danger is the number of dangerous events occurring during flight, M flights is the total number of flights; E efficiency represents the mission execution efficiency T planned is the time to complete the planned mission, T actual is the actual time to complete the mission; E stability represents the trajectory stability of the aircraft and are the velocity vectors of the nth and (n - 1)th predicted trajectory points respectively; E robustness represents the robustness of the model F failures is the number of mission failures, F total is the total number of missions

[0031] Preferably, the evolutionary learning model is a deep neural network, a convolutional neural network, an LSTM or a reinforcement learning model

[0032] On the other hand, the present invention also provides a computer-readable storage medium storing one or more programs, the one or more programs including instructions which, when executed by a computing device, cause the computing device to execute the above-mentioned intelligent evolutionary learning method for the operating trajectory of an unmanned aerial vehicle based on virtual simulation

[0033] On the other hand, the present invention also provides an electronic device, including one or more processors, one or more memories and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing the above-mentioned intelligent evolutionary learning method for the operating trajectory of an unmanned aerial vehicle based on virtual simulation

[0034] Beneficial effects: The intelligent evolutionary learning method for the operating trajectory of an unmanned aerial vehicle based on virtual simulation proposed by the present invention can effectively improve the accuracy and robustness of the flight trajectory prediction of the aircraft

[0035] First of all, through the dynamic evaluation and cross-evolution mechanism of the data set, the composition of the data set can be automatically adjusted according to historical data and real-time simulation feedback, making the model training more flexible and improving the representativeness of the training data; especially through the sampling and weighting of low-quality samples, overfitting can be effectively prevented, and the generalization ability of the model can be improved, enabling it to show high adaptability under different environmental and task conditions

[0036] Secondly, by fusing the real-time generated simulation data with the initial data set and gradually adding them to the data set after sorting according to the evaluation criteria, it is ensured that the data set can reflect the dynamic changes of the environment and the diversity of flight missions during the continuous evolution process. This process mimics the cross-evolution mechanism of organisms, which helps the model extract useful features from data of different qualities, thereby improving the robustness of the model in complex tasks. Especially when facing emergencies such as dynamic obstacles, meteorological changes, and equipment failures, the model can maintain strong adaptability.

[0037] Finally, through continuous iterative learning and optimization, the method of the present invention not only improves the accuracy and efficiency of aircraft trajectory prediction, but also effectively reduces the training cost. Especially in complex flight missions and changing environments, it can ensure the stability and safety of flight missions, and has broad application prospects and commercial value. Description of the Drawings

[0038] Figure 1 It is the flowchart of the method of the embodiment of the present invention. Detailed Embodiment

[0039] The present invention proposes an intelligent evolutionary learning method for the flight trajectory of an aircraft based on virtual simulation. The method includes: establishing an aircraft trajectory library composed of several trajectory segments, and adding high-quality samples and low-quality samples, and adding high-quality samples and some low-quality samples extracted according to set rules to the sample set, and dividing the sample set into a training set and a test set according to a ratio; defining a loss function considering the comprehensive cost, and determining the comprehensive evaluation index through system simulation; training the evolutionary learning model and controlling the aircraft in the simulation environment to output new aircraft trajectory samples. If the requirements of the system simulation evaluation index are met, the process ends, otherwise, the new trajectory generated by the simulation is combined with the previous trajectories to form a new trajectory library, and returns for re-iterative optimization. This method can effectively improve the efficiency and accuracy of aircraft trajectory optimization, while reducing the training cost, and is applicable to the trajectory optimization of unmanned aircraft in complex flight mission scenarios.

[0040] As Figure 1 shown, the intelligent evolutionary learning method for the flight trajectory of an unmanned aircraft based on virtual simulation of the present invention includes:

[0041] (1) Establish an aircraft trajectory library including several trajectory segments, and divide all trajectory segments into two categories: high-quality and low-quality;

[0042] (2) Select trajectory segments from the trajectory library to obtain a trajectory segment sample set for training the evolutionary learning model;

[0043] (3) Train the evolutionary learning model based on the sample set selected in step (2);

[0044] (4) Put the trained evolutionary learning model into the simulation environment for simulation testing. If the aircraft satisfies the requirements of the preset simulation comprehensive evaluation index when performing the flight mission according to the predicted trajectory output by the trained evolutionary learning model, the algorithm ends, and the trained evolutionary learning model is solidified; otherwise, return to step (1), and supplement the predicted trajectory output by the trained evolutionary learning model into the trajectory library to form a new trajectory library.

[0045] Specifically, the specific implementation method of step (1) is as follows:

[0046] Read the trajectory library of the aircraft, and use the sliding window method to divide the continuous trajectory points into multiple trajectory segments with a length of K. The value of K is set by the user, and the total number of trajectory segments N is counted. total . For each trajectory segment, score it according to the comprehensive evaluation criteria. The evaluation dimensions include key indicators such as trajectory accuracy, smoothness, energy consumption, and stability. Sort the trajectory segments in descending order according to the scoring results, and divide them into high-quality and low-quality categories as needed for the experience initialization and policy optimization reference of the reinforcement learning model.

[0047] The comprehensive evaluation criteria for scoring the trajectory segments are as follows:

[0048] Score(τ k ) = ω1·P k +ω2·S k +ω3·E k +ω4·R k

[0049] Where: P k is the accuracy score of the k-th trajectory segment; S k is the smoothness score of the k-th trajectory segment; e k is the energy consumption score of the k-th trajectory segment; R k is the stability score of the k-th trajectory segment; ω1, ω2, ω3, and ω4 are the weights of the corresponding indicators respectively.

[0050] Divide the trajectory segments in the trajectory library into high-quality samples and low-quality samples by setting the ratio P of high-quality samples to low-quality samples, which is convenient for subsequent sampling operations on low-quality samples. The calculation of the ratio P of high-quality samples to low-quality samples depends on the sample standard deviation and mean, and introduces clustering analysis and dynamic adjustment of errors:

[0051]

[0052] Where: N total is the total number of trajectory segments in the trajectory library; i is the sample number. In the historical trajectory library of the trajectory library, all samples have their own numbers, and the range is [1, N total; error(i) is the model prediction error of the i-th sample, and the model prediction error of the i-th sample is obtained by calculating the difference between the model prediction trajectory and the actual trajectory of this sample; error max is the maximum prediction error among all samples, that is, after traversing all samples in the trajectory library, the maximum value is selected from the calculated respective prediction errors; λ is the sensitivity to control the dynamic adjustment of the clustering factor to the error; ω cluster (i) is the weight of the cluster (high-quality trajectory segment or low-quality trajectory segment) to which the i-th sample belongs, and the value is determined by empirical setting or simulation evaluation results.

[0053] Specifically, the number of samples N in the trajectory segment sample set in step (2) sample refers to the number of samples required for model training, which is selected from the total number of historical trajectory segments N in the trajectory library total , and the sampling strategy is: all high-quality samples and part of the low-quality samples extracted in descending order according to P sample .

[0054] N sample The specific value of depends on the acceptable maximum training time (T max ) and the expected relative improvement ratio of model performance (R exp ):

[0055]

[0056] where: T max is the acceptable maximum training time, in seconds; Step is the average training time per sample (unit: seconds / sample); R exp : the expected relative improvement ratio of model performance (range in (0, 10%]), and the specific value is set by the user.

[0057] The sampling strategy for low-quality samples is based on reinforcement learning, and reinforcement learning methods such as Q-learning are used to calculate the extraction probability of each low-quality sample:

[0058]

[0059] where: P sample (j) is the extraction probability of the j-th low-quality trajectory segment; Q(j) is the Q value of the j-th low-quality sample, which is dynamically calculated by the reinforcement learning model. The larger the Q value, the higher the probability of the sample being extracted.

[0060] Specifically, in step (3), the sample set is dynamically divided to obtain the training set and the test set, and the calculation formula for the ratio of the number of trajectory segments in the training set and the test set is:

[0061]

[0062] Where: R init is the preset initial training set and test set ratio; α is the adjustment parameter that controls the sensitivity of ratio adjustment; N epoch is the current iteration number of training.

[0063] Furthermore, the evolutionary learning model is trained using a loss function that considers the comprehensive cost:

[0064] L = α1·L position + α2·L energy + α3·L time + α4·L smoot ++ α5·L collision

[0065] Where:

[0066] ● α1, α2, α3, α4, α5 are the weight coefficients of each index;

[0067] ·L position is the trajectory position error loss, which represents the sum of the differences between the actual positions and the predicted positions of all trajectory points in the trajectory segment. The calculation formula is:

[0068]

[0069] In the formula, p n and are the actual position and the predicted position of the nth trajectory point, and N is the number of trajectory points in the trajectory segment.

[0070] ●L energy is the energy consumption loss, which represents the product of the energy consumption of all trajectory points in the trajectory segment and the unit energy cost. The calculation formula is:

[0071]

[0072] In the formula, E n is the energy consumption of the nth trajectory point (unit: joule or kilowatt-hour), and C energy is the cost of unit energy.

[0073] ●L time is the time consumption loss, which represents the product of the time required to complete on this trajectory segment and the unit time cost. The calculation formula is:

[0074] L time = T total ·C time

[0075] In the formula, T totalis the total time required for the aircraft to complete a flight mission (unit: seconds or hours), C time is the cost per unit time.

[0076] ●L smooth is the path smoothness loss, which measures the smoothness of the trajectory and excessive turning. A more complex path may lead to higher energy consumption and unstable execution. The calculation formula is:

[0077]

[0078] where v n is the velocity vector of the nth trajectory point, and v n+1 is the velocity vector of the (n - 1)th trajectory point. This loss measures the velocity difference between adjacent trajectory points and is used to optimize the smoothness of the path.

[0079] ·L collision is the collision risk loss, indicating whether the trajectory is likely to collide with obstacles or other aircraft.

[0080] The calculation formula for this loss is:

[0081]

[0082] where, P collision (p n ) is the collision probability of the nth trajectory point with the obstacle. If the trajectory point is close to the obstacle, the collision probability is high, and the corresponding loss increases.

[0083] Furthermore, determine the loss function L considering the comprehensive cost, select an appropriate model structure, such as a deep neural network, convolutional neural network, LSTM, reinforcement learning, etc. for training. During the training process, through optimization algorithms, such as Adam, SGD, evolutionary strategies, etc., continuously update the model parameters to ensure that the model can accurately predict the aircraft trajectory and achieve the task objectives. At the same time, to improve the efficiency and accuracy of training, combined with historical data and the dynamic feedback of the simulation environment, adopt online learning and transfer learning strategies to accelerate the convergence of the model. After completing the model training, bring the trained model into the simulation environment for testing. In the simulation environment, the trained evolutionary learning model will guide the movement trajectory of the aircraft in real time through the control system of the aircraft. The simulation environment needs to include the physical model, dynamic model of the aircraft, as well as dynamic obstacles and environmental changes. The aircraft executes the flight mission according to the predicted trajectory output by the model, and evaluates the simulation effect through the comprehensive evaluation index E set by the system simulation.

[0084] Specifically, the method for determining the simulation comprehensive evaluation index E is:

[0085] E = β1·E accuracy +β2·Esafety +β3·E efficiency +β4·E stability +β5·E robustness

[0086] Wherein:

[0087] ·β1, β2, β3, β4, and β5 are the weight coefficients of each index;

[0088] ●E accuracy represents the accuracy index of the trajectory, evaluating the matching degree between the predicted trajectory of the model and the actual trajectory. The calculation method of this index is:

[0089]

[0090] In the formula, M correct is the number of correctly predicted trajectory points, and M total is the total number of trajectory points. This index can reflect the accuracy of the trajectory optimization model.

[0091] ●E safety represents the flight safety index, evaluating whether dangerous events occur during the flight. The calculation method of this index is:

[0092]

[0093] In the formula, M danger is the number of dangerous events that occur during the flight, and M fl is the total number of flights. This index is used to reflect flight safety. The fewer dangerous events, the higher the safety.

[0094] ●E efficiency represents the task execution efficiency, evaluating the timeliness of task completion, considering the difference between the planned completion time and the actual completion time. Its calculation formula is:

[0095]

[0096] In the formula, T planned is the time to complete the task as planned, and T actual is the actual time to complete the task. This index reflects the efficiency of task execution. The closer the value is to 1, the more efficient the task is.

[0097] ●E stability represents the trajectory stability of the aircraft, evaluating the fluctuation of the trajectory during the flight. It can be measured by the smoothness of the trajectory, speed changes, etc. The calculation method is:

[0098]

[0099] In the formula, and They are the velocity vectors of the nth and (n - 1)th predicted trajectory points respectively. This index reflects the stability of the trajectory. The smaller the trajectory change, the higher the stability.

[0100] ·E robustness It represents the robustness of the model and evaluates the stability of the model when facing different environmental changes, task difficulties or noise interferences. This index considers the adaptability of the model to the environment and can be measured by the task failure rate:

[0101]

[0102] In the formula, F failures is the number of task failures, and F total is the total number of tasks. This index reflects the performance of the model in a complex environment. The lower the failure rate, the stronger the robustness.

[0103] To make the content of the present invention easier to be clearly understood, the following further describes the present invention in detail with specific embodiments.

[0104] Suppose a drone performs a reconnaissance mission. It needs to fly in a complex environment and optimize its flight trajectory through evolutionary learning to improve the efficiency and safety of task completion.

[0105] 1. Trajectory library reading and sorting

[0106] For the N total = 2100 trajectory segments in the trajectory library, score them according to the comprehensive evaluation criteria, sort them from high to low, and classify them according to the ratio P of high-quality trajectory segments and low-quality trajectory segments.

[0107] Suppose the following parameters are known:

[0108] ●N low = 300;

[0109] ●error(i) = 0.01 (i = 0, 1, 2, 3...)

[0110] ●error max = 0.01;

[0111] ●λ = 0.1;

[0112] ●ω cluster (i) = 1 (i = 0, 1, 2, 3...)

[0113] Then:

[0114]

[0115] 2. Determine the number of samples N used to train the evolutionary learning model sample, set the sampling method P sample Extract low-quality samples and add them to the set, and set the ratio R of the training set and the test set tr .

[0116] Assume the following parameters are known:

[0117] ● T max The maximum acceptable training time is 100s;

[0118] ● Δ effect The expected improvement ratio of the model effect is 0.05;

[0119] ● Step = 0.1 second / sample;

[0120] ● R exp = 0.1%;

[0121] ● N total = 2100;

[0122] According to the formula Calculate the number of samples N required for training sample = 2100.

[0123] Assume the following parameters are known:

[0124] ● Q(i) = 1 (i = 0, 1, 2, 3...);

[0125]

[0126] Assume the following parameters are known:

[0127] ● R init = 0.142;

[0128] ● α = 0.1;

[0129]

[0130] 3. Determine the loss function

[0131] At each step of model training, a loss function will be obtained. Here, taking one training as an example, the calculation process of the loss function is shown. Assume the following parameters are known:

[0132] ● α1, α2, α3, α4, α5 = 1

[0133] ● N = 500;

[0134] ● For simplicity of calculation, assume p n and The difference is always 0.1;

[0135] ● En = 0.2 (n = 0, 1, 2, 3...);

[0136] ● C energy = 5;

[0137] ● T total = 10;

[0138] ● C time = 2;

[0139] ●

[0140] ● P collision = 0.1

[0141] From this, it can be obtained that:

[0142]

[0143] L time = T total · C time = 10 × 2 = 20

[0144]

[0145] Therefore:

[0146] L = 50 + 500 + 20 + 0 + 50 = 620.

[0147] 4. Comprehensive evaluation indicators of system simulation

[0148] Assume the following parameters are known:

[0149] ● β1, β2, β3, β4, β5 = 1;

[0150] · M correct = 400;

[0151] ● N = 500;

[0152] ● M danger = 0;

[0153] ● M flights = 100;

[0154] · T planned = 100;

[0155] ● T actual = 80;

[0156] ●

[0157] ● F failures = 0;

[0158] ● Ftotal = 100;

[0159] From this, it can be obtained that:

[0160]

[0161] Therefore:

[0162] E = 0.80 + 1 + 0.80 + 0.90 + 1 = 4.50.

[0163] 5. Training of the evolutionary learning model and trajectory output

[0164] Based on the above loss function L considering comprehensive costs, select an appropriate model structure, such as a deep neural network, convolutional neural network, LSTM, reinforcement learning, etc. for training. During the training process, through optimization algorithms, such as Adam, SGD, evolutionary strategies, etc., continuously update the model parameters to ensure that the model can accurately predict the flight trajectory of the aircraft and achieve the mission objectives. At the same time, in order to improve the efficiency and accuracy of training, combined with historical data and the dynamic feedback of the simulation environment, adopt online learning and transfer learning strategies to accelerate the convergence of the model.

[0165] In the simulation environment, the trained evolutionary learning model will guide the movement trajectory of the aircraft in real time through the control system of the aircraft. The simulation environment needs to include the physical model, dynamic model of the aircraft, as well as dynamic obstacles and environmental changes. The aircraft executes the flight mission according to the predicted trajectory output by the model, and evaluates the simulation effect through the comprehensive evaluation index E of the system simulation.

[0166] 6. Training of the evolutionary learning model and trajectory output

[0167] Output a new aircraft trajectory sample. If it meets the requirements of the comprehensive evaluation index of the system simulation, then end; otherwise, combine the new trajectory generated by the simulation with the previous old trajectory to form a new trajectory library, and repeat the above process.

[0168] Based on the same technical solution, the present invention also provides a computer-readable storage medium storing one or more programs, the one or more programs including instructions, characterized in that when the instructions are executed by a computing device, the computing device executes the intelligent evolutionary learning method for the flight trajectory of an unmanned aircraft based on virtual simulation as described above.

[0169] Based on the same technical solution, the present invention also provides an electronic system, including one or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing the intelligent evolutionary learning method for the flight trajectory of an unmanned aircraft based on virtual simulation as described above.

[0170] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) that contain computer-usable program code.

[0171] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for realizing the functions specified in Figure 1 one or more of the flows Figure 1 or blocks or the combination of blocks.

[0172] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing devices to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that realizes the functions specified in Figure 1 one or more of the flows Figure 1 or blocks or the combination of blocks.

[0173] These computer program instructions can also be loaded onto a computer or other programmable data processing devices, such that a series of operation steps are executed on the computer or other programmable devices to generate a computer-implemented process, so that the instructions executed on the computer or other programmable devices provide steps for realizing the functions specified in Figure 1 one or more of the flows Figure 1 or blocks or the combination of blocks.

[0174] Obviously, the above embodiments are merely examples given for clear illustration and are not limitations on the usage. For those skilled in the art, other different forms of changes or variations can be made based on the above description. It is not necessary and impossible to exhaustively list all usage manners here. And the obvious changes or variations derived therefrom are still within the protection scope of the present invention.

Claims

1. An intelligent evolutionary learning method for the operating trajectory of an unmanned aerial vehicle based on virtual simulation, characterized in that, The method includes: (1) Establish an aircraft trajectory library including several trajectory segments, and divide all trajectory segments into two categories: high-quality and low-quality; (2) Select trajectory segments from the trajectory library to obtain a trajectory segment sample set for training an evolutionary learning model; (3) Based on the sample set selected in step (2), train the evolutionary learning model; (4) Put the trained evolutionary learning model into a simulation environment for simulation testing. If the aircraft satisfies the requirements of the preset simulation comprehensive evaluation index when performing a flight mission according to the predicted trajectory output by the trained evolutionary learning model, the algorithm ends, and the trained evolutionary learning model is solidified; otherwise, return to step (1), and supplement the predicted trajectory output by the trained evolutionary learning model into the trajectory library to form a new trajectory library.

2. The method according to claim 1, wherein In step (1), the calculation formula for the ratio P between the number of high-quality trajectory segments and the number of low-quality trajectory segments is: Where: N total is the total number of trajectory segments in the trajectory library; error(i) is the prediction error of the evolutionary learning model for the i-th trajectory segment in the trajectory; error max is the maximum prediction error among all trajectory segments in the trajectory; λ is the sensitivity parameter; ω cluster (i) is the weight value corresponding to whether the i-th trajectory segment in the trajectory belongs to a high-quality trajectory segment or a low-quality trajectory segment.

3. The method according to claim 1, wherein The number of samples N in the sample set used to train the evolutionary learning model in step (2) sample The calculation formula is as follows: Where: T max is the maximum acceptable training time; Step is the average training time per trajectory segment sample; R exp is the expected relative improvement ratio of model performance.

4. The method according to claim 1, wherein The method for selecting trajectory segments from the trajectory library in step (2) includes: (2.1) Select all high-quality trajectory segments in the trajectory library into the sample set; (2.2) Extract a certain number of low-quality trajectory segments in descending order of extraction probability to meet the sample quantity requirements of the sample set; Among them, the calculation formula for the extraction probability includes: Where: P sample (j) is the probability that the j-th low-quality trajectory segment is extracted; Q(j) is the Q value of the j-th low-quality sample; N low is the total number of low-quality trajectory segments.

5. The method according to claim 1, wherein In step (3), training the evolutionary learning model further includes: dynamically dividing the sample set to obtain a training set and a test set, where the calculation formula for the ratio R of the number of trajectory segments in the training set and the test set is: Where: R init is the preset initial training set and test set ratio; α is the adjustment parameter; N epochs is the current iteration number of training.

6. The method according to claim 1, wherein The expression of the loss function for training the evolutionary learning model in step (3) is: L = α1·L position + α2·L energy + α3·L time + α4·L smooth ++ α5·L collision where: α1, α2, α3, α4, α5 are weight coefficients; L position is the trajectory position error loss, p n is the actual position of the nth trajectory point in the trajectory segment sample, is the predicted position of the nth trajectory point, and N is the number of trajectory points in the trajectory segment sample; L energy is the energy consumption loss, E n is the energy consumption of the nth trajectory point, and C energy is the cost per unit of energy; L time is the time consumption loss, L time = T total ·C time where T total is the total time required for the aircraft to complete a flight mission, and C time is the cost per unit of time; L smoo is the path smoothness loss, in which v n is the velocity vector of the nth trajectory point, and v n+1 is the velocity vector of the (n - 1)th trajectory point; L collision is the collision risk loss, is the collision probability between the nth trajectory point and the obstacle.

7. The method according to claim 1, wherein The calculation formula for the simulation comprehensive evaluation index in step (4) is as follows: E = β1·E accuracy + β2·E safety + β3·E efficiency + β4·E stability + β5·E robustness where: β1, β2, β3, β4, β5 are weight coefficients; e accuracy represents the accuracy index of the trajectory, M correct is the number of correctly predicted trajectory points, M total is the total number of trajectory points; E safety represents the flight safety index, M danger is the number of dangerous events occurring during the flight, M fligh is the total number of flights; E efficiency represents the mission execution efficiency, T planned is the time to complete the planned mission, T actual is the actual time to complete the mission; E stability represents the stability of the aircraft trajectory, and are the velocity vectors of the nth and (n - 1)th predicted trajectory points respectively; E robustness represents the robustness of the model, F failures is the number of mission failures, F total is the total number of missions.

8. The method according to claim 1, wherein The evolutionary learning model is a deep neural network, a convolutional neural network, an LSTM, or a reinforcement learning model.

9. A computer-readable storage medium storing one or more programs, the one or more programs including instructions, characterized in that, When executed by a computing device, the instructions cause the computing device to execute the method according to any one of claims 1 to 8.

10. An electronic device, characterized in that, Comprising one or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and configured to be executed by the one or more processors, and the one or more programs include instructions for executing the method according to any one of claims 1 to 8.

Images