An unmanned aerial vehicle simulation flight intelligent control method and system for college teaching

Abstract

This application discloses an intelligent control method and system for simulated flight of unmanned aerial vehicles (UAVs) for university teaching, including: a model construction step, generating a scene-adapted model; and collecting students' control data during flight, and identifying students' personalized control characteristic parameters through an online parameter identification algorithm; a safety construction step, constructing a physical safety envelope and a skill safety envelope; an intent recognition step; an intervention step; and a teaching optimization step, collecting the control commands and flight response data, identifying error types based on a misoperation feature database, and generating teaching prompt information. Thus, this application embodiment constructs a multi-rotor UAV dynamics model and adapts the parameters to university teaching scenarios (such as playgrounds or airspace-restricted environments), configuring and controlling the UAV's flight based on these parameters. This allows the control system to correct flight attitude, throttle, and heading control quantities in real time according to student operations, and automatically identify misoperations and provide safety interventions.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) teaching technology, and in particular to an intelligent control method and system for UAV simulated flight for university teaching. Background Technology

[0002] In university drone training, real drones are typically used as the teaching medium. Students, under the guidance of instructors, conduct direct flight control training in open outdoor areas or dedicated airspaces. The training covers basic skills such as takeoff, hovering, flight path planning, attitude correction, and emergency response. However, this existing model has the following shortcomings: First, real drones and related equipment are expensive, and the margin for error in flight operations is extremely low. Crashes, collisions, or other operational errors can easily lead to complete destruction of the drone, resulting in high equipment wear and maintenance costs. Second, real drone training carries high safety risks. Operational errors can not only damage equipment but also pose threats to personnel and surrounding facilities, forcing teachers and students to be overly tense during training. Third, real drone flight has stringent requirements for site conditions, needing to meet multiple constraints such as airspace clearance and airspace approval. Universities generally lack dedicated flight sites, and existing sites are insufficient in terms of area, opening hours, and frequency of use to meet the teaching needs of high-frequency, large-scale flights. Summary of the Invention

[0003] This application aims to address at least one of the technical problems existing in the prior art. To this end, one objective of this application is to propose an intelligent control method and system for UAV simulated flight in higher education, enabling the construction of an effective UAV teaching and guidance control approach based on preset scenarios.

[0004] Firstly, according to an embodiment of this application, a method for intelligent control of unmanned aerial vehicle (UAV) simulated flight for university teaching includes: a model construction step, which involves acquiring type information of the current teaching scenario, retrieving corresponding environmental feature parameters based on the type information, resetting parameters of a preset UAV dynamics model, and generating a scenario-adaptive model; and collecting student control data during flight, identifying personalized control feature parameters of students through an online parameter identification algorithm, embedding the personalized control feature parameters into the scenario-adaptive model, and generating a dedicated simulation model; a safety construction step, which involves constructing a physical safety envelope and a skill safety envelope, wherein the physical safety envelope is set based on the physical limit parameters of the UAV, and the skill safety envelope is constructed based on the statistical distribution of students' historical normal operation data; and intent recognition. The process involves several steps: First, real-time acquisition of students' control command sequences. These sequences are then input into an intent prediction network based on a long short-term memory network, which outputs control intent categories, including routine training intents, high-maneuverability intents, and suspected misoperation intents. Second, an intervention step involves implementing a tiered intervention strategy based on the position of the control commands relative to the physical safety envelope, the skill safety envelope, and the control intent category. Third, a teaching optimization step involves collecting control commands and flight response data, identifying error types based on a misoperation feature library, generating teaching prompts, and decomposing the flight mission into basic control units. Features of each unit are extracted and input into a defect classifier to output defect type labels. Personalized compensation training tasks are then dynamically generated from a compensation training library based on these defect type labels.

[0005] The effect is that, by constructing a dynamic model of a multi-rotor UAV and adapting the parameters to university teaching scenarios (such as playgrounds or restricted airspace environments), the UAV's flight can be controlled based on these parameters. This allows the control system to correct the flight attitude, throttle, and heading control in real time according to student operations, and to automatically identify misoperations and provide safety interventions. Therefore, the accident rate in practical training can be significantly reduced, and the stability of novice operators can be improved.

[0006] Furthermore, the tiered intervention strategy includes: when the control command exceeds the skill safety envelope and the control intention category is a regular training intention, performing response passivation processing; when the control command exceeds the skill safety envelope and the control intention category is a suspected misoperation intention, performing active correction intervention; when the control command exceeds the physical safety envelope, performing safety locking; then, monitoring whether the flight response data output by the dedicated simulation model exceeds the preset safety envelope threshold, if it exceeds, activating shadow mode, weighted fusing the real-time control command with the preset standard safety trajectory command, and generating a correction control quantity to replace the real-time control command.

[0007] Furthermore, the response passivation processing includes: linearly increasing the damping coefficient according to the degree to which the control command exceeds the skill safety envelope, so that the amplitude of the angular acceleration response of the flight simulator is reduced proportionally, and the reduction ratio is positively correlated with the degree of exceedance; the active correction intervention includes: vector compression of the control command towards the skill safety envelope boundary, and the compressed command replacing the original command input to the dedicated simulation model; the safety locking includes: smoothly switching the control of the dedicated simulation model to an automatic hovering state.

[0008] Furthermore, the environmental characteristic parameters include: for the playground environment, a wind disturbance matrix is ​​introduced; for the airspace-constrained environment, a virtual potential field algorithm is superimposed on the UAV dynamics model to generate a reverse damping force when the aircraft approaches a preset electronic fence; the dynamic equations of the position loop are updated as follows:

[0009] ;

[0010] Where m is the mass and P is the position vector. To simulate the acceleration of an aircraft, This is the acceleration vector caused by wind disturbance. The reverse damping force is T, and the total thrust of the propeller is T. Let t be the rotation matrix from the body coordinate system to the ground coordinate system, and t be the time.

[0011] Furthermore, the online parameter identification algorithm employs a recursive least squares algorithm to continuously update the personalized control feature parameters using a sliding time window, the width of which is adaptively adjusted according to the student's training frequency.

[0012] Furthermore, the skill safety envelope is constructed based on the statistical distribution of students' historical normal operation data, including: a Gaussian process regression model with control commands as input and the probability of reasonable operation as output. The boundary of the skill safety envelope corresponds to a preset reasonableness probability threshold. :

[0013] In the formula, These are new control commands. This indicates that the Gaussian process model corresponds to the control commands. The predicted mean, This is a preset probability threshold for reasonableness. Furthermore, the category of manipulation intent is obtained through an intent prediction network, which is a time-series classification model based on a Long Short-Term Memory (LSTM) network. Its input is a continuous sequence of lever movements within a fixed time window, and its output is a probability distribution of regular training intent, high-maneuvering intent, or suspected misoperation intent. The intent prediction network is a time-series classification model built on an LSTM network, and its input is a fixed time window. Continuous rod quantity sequence within The output is the probability distribution of the manipulation intention category C:

[0014] Among them, the probability of the intent category :

[0015] ;

[0016] In the formula, It is the weight matrix of the output layer. It is the bias vector of the output layer. It is the hidden state of the last time step; the manipulation intent category is... .

[0017] Furthermore, in the shadow mode, the weighting coefficients of the weighted fusion are dynamically adjusted according to the degree to which the flight response data exceeds the safety envelope threshold. The greater the degree of exceedance, the higher the weight of the standard safety trajectory command. This includes: setting the standard safety trajectory command as... Weighted fusion corrected control quantity for:

[0018] ;

[0019] Among them, the weighting coefficient The system dynamically adjusts based on the degree to which flight response data exceeds the safety envelope threshold.

[0020] Furthermore, the error feature library includes attitude confusion features, throttle loss features, and visual line-of-sight panic features; the step of identifying error types specifically includes: extracting the stick change rate of the control command and the attitude lag angle in the flight response data, and inputting the extracted features into a pre-trained classifier to output the corresponding error type.

[0021] Furthermore, the defect classifier is a support vector machine or a multilayer perceptron, and the features include time-domain features, frequency-domain features, and phase features; the time-domain features include rise time, overshoot, and steady-state error; the frequency-domain features include the energy distribution of the lever power spectral density within a preset frequency band; and the phase feature is the peak delay time of the cross-correlation function between the input lever quantity and the output attitude angle. The compensation training library contains multiple basic compensation training units, each associated with a corresponding defect type label and training effect weights. The dynamic combination to generate personalized compensation training tasks includes selecting candidate units based on defect type labels and performing combination optimization with the goal of maximizing the overall training effect: the objective function is defined as:

[0022] ;

[0023] The optimization goal is Set constraints ;

[0024] in, Let i be the time consumed for the i-th training unit. To the maximum allowed training duration, It is a decision variable. These are training effect weights. These are the defect labels of the training units. It is a similarity function. Let be the time taken for the i-th training unit.

[0025] Secondly, according to an embodiment of this application, an intelligent control system for drone simulation flight for university teaching includes a processor and a storage device. The storage device stores a computer program for implementing the aforementioned intelligent control method for drone simulation flight for university teaching. The processor performs read and write operations on the storage device.

[0026] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0027] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0028] Figure 1 This is a simplified schematic diagram illustrating the steps of an intelligent control method for drone simulation flight in college teaching, according to an embodiment of this application. Detailed Implementation

[0029] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0030] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0031] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0032] An intelligent control method for unmanned aerial vehicle (UAV) simulated flight for university teaching, according to an embodiment of this application, includes:

[0033] The model building steps involve obtaining the type information of the current teaching scenario, retrieving the corresponding environmental feature parameters based on the type information, resetting the parameters of the preset UAV dynamics model, and generating a scenario-adaptive model.

[0034] Specifically, this embodiment dynamically injects environmental adaptability and student-specific characteristics into a standard UAV dynamics model, thereby generating a customized simulation model that highly matches the student's current skill level and the teaching scenario. This process is not a simple parameter replacement, but a continuous process that integrates environmental modeling and online identification.

[0035] For example, a general dynamic model for a UAV is defined, using a simplified model of a quadcopter's attitude loop (inner loop) and position loop (outer loop) in the body coordinate system. The standard dynamic model can be represented as:

[0036] Posture ring: ;

[0037] Position ring:

[0038] In the formula, Euler angles (such as roll, pitch, and yaw). Let m be the angular velocity of the body, m be the mass, and P be the position vector. To simulate the acceleration of an aircraft, ω is the angular velocity, g is the acceleration due to gravity. Let T be the rotation matrix from the body coordinate system to the ground coordinate system, and let T be the total thrust of the propeller.

[0039] Then, the parameters of the preset UAV dynamics model are reset to generate a scene-adaptive model. In this embodiment, after obtaining the type information of the current teaching scene (such as indoor, playground, or airspace-restricted environment), the corresponding set of environmental feature parameters is retrieved. It is worth noting that since the training site is fixed, the aforementioned set of feature parameters can store the pre-collected parameters.

[0040] For example, for indoor environments, the ground effect coefficient can be increased, the position loop control gain can be reduced, or visual positioning drift noise can be introduced.

[0041] For example, a time-varying wind disturbance matrix can be introduced for the playground environment. Applying this to the position loop yields the acceleration vector caused by wind disturbance. :

[0042] ;

[0043] Where t is time, and the wind disturbance matrix is... It can be further modeled as a wind field model such as Dryden's (an atmospheric turbulence model based on stochastic process theory), whose parameters (such as turbulence intensity and scale) are preset according to the scenario, thereby simulating the acceleration of the aircraft. Wind disturbance components were superimposed on the original dynamics.

[0044] Furthermore, a virtual potential field algorithm is superimposed on the UAV dynamics model to generate a reverse damping force when the aircraft approaches a preset electronic fence. In this example, a virtual potential field algorithm is superimposed on the dynamics model, which generates a position-dependent reverse damping force. This is used to provide smooth constraint when an aircraft approaches a pre-defined electronic fence. The force is defined as a potential field function. The negative gradient.

[0045] The potential field function can be designed as, for example, a barrier function that tends to infinity at the boundary, or a repulsive field that increases linearly upon entering the warning zone. Then, the dynamic equations of the position loop are updated as follows:

[0046]

[0047] Therefore, environmental characteristic parameters are integrated into the dynamic model to generate a scene-adaptive model. This model serves as the basis for operations, and its parameters are determined by the scene type and remain fixed within a single flight mission.

[0048] Furthermore, the system collects students' control data during flight, uses an online parameter identification algorithm to invert and identify students' personalized control feature parameters, and embeds these personalized control feature parameters into the scenario adaptation model to generate a unique simulation model.

[0049] Specifically, let the student's manipulation data at time t be... For example, the joystick's lever position; the actual response data of the aircraft in the simulation environment is... Examples include attitude angle and position. A mapping relationship for student-drone interaction is constructed, which is represented by a vector with unknown parameters. The model f is represented by .

[0050] ;

[0051] in, These are personalized control feature parameters used to characterize the user's operational fingerprint.

[0052] In this embodiment, the online parameter identification algorithm employs a recursive least squares algorithm, continuously updating the personalized control feature parameters using a sliding time window. The width of the sliding time window is adaptively adjusted based on the student's training frequency. Specifically, in order to identify... An online parameter identification algorithm, specifically the recursive least squares (RLS) algorithm, was employed. This allows for the identification of parameters based on new control data. , Input, continuous optimization This minimizes the error between the model's predicted output and the actual output. In some embodiments, the width of the sliding time window during the update process can be adaptively adjusted according to the student's training frequency. For students with high training frequency, parameter changes may be more drastic, requiring a smaller time window width to track changes more quickly; for students with low training frequency, a larger time window width is used to maintain the stability of parameter estimation. Through the RLS algorithm, the system can continuously and in real time retrieve the student's current personalized manipulation feature parameters.

[0053] Finally, based on Real-time updated parameter vector This is embedded into the previously generated scene adaptation model, thereby constructing a customized simulation model that dynamically reflects the student's current control style. This model serves as both the environment for the student's simulated flight and the basis for all subsequent safety assessments and instructional interventions.

[0054] The safety construction steps involve constructing a physical safety envelope and a skill safety envelope S. The physical safety envelope is based on the physical limit parameters of the UAV, and the specifics are not elaborated here. The skill safety envelope S is constructed based on the statistical distribution of students' historical normal operation data, including a Gaussian process regression model with control commands as input and the probability of operation rationality as output. The boundary of the skill safety envelope corresponds to a preset rationality probability threshold.

[0055] In detail, the physical safety envelope is a static, absolute safety boundary based on the physical limit parameters of the UAV. Thresholds can be set for key state quantities of the aircraft, such as attitude angle, angular velocity, airspeed, and altitude; further details will not be elaborated here.

[0056] For skill safety envelope This is a dynamic, personalized safety boundary constructed based on the statistical distribution of students' historical normal operation data; in this embodiment, a Gaussian process regression model is used to learn from control commands. To the probability of operational rationality The model not only provides the predicted values, but also the uncertainty of those predictions.

[0057] For example, when building the model, the input is the control commands from historical normal operation data. The output is the corresponding rationality tag. (Representing "normal"). A function is learned through Gaussian process regression. , making The core of a Gaussian process is to define a covariance function (kernel function) to measure the similarity between two control commands, so as to calculate the predicted distribution of their corresponding probability of reasonableness through Gaussian process regression.

[0058] ;

[0059]

[0060] in, These are new control commands. This indicates that the Gaussian process model corresponds to the control commands. The predicted mean, Indicates test point With all training points The covariance vector between training points, where K represents the covariance matrix between training points and I represents the identity matrix. y represents the noise variance, and y represents the output vector corresponding to the training sample.

[0061] In this way, the boundary of the skill safety envelope is obtained. :

[0062]

[0063] in, In this example, It is a preset reasonableness probability threshold (e.g., 0.95) used to define the boundary of the skill safety envelope; when the student's control instruction u is inside the envelope, it is considered a safe operation that conforms to his / her historical skill level; when the instruction exceeds the envelope, it means that the operation exceeds the student's current ability range and has a high risk.

[0064] The intent recognition step involves real-time acquisition of the student's control command sequence. This sequence is then input into an intent prediction network based on a Long Short-Term Memory (LSTM) network, which outputs a control intent category. This category includes regular training intent, high-mobility intent, and suspected misoperation intent. In this embodiment, the control intent category is obtained through the intent prediction network, a time-series classification model based on LSM. Its input is a continuous sequence of lever movements within a fixed time window, and its output is a probability distribution of regular training intent, high-mobility intent, or suspected misoperation intent. Based on the constructed physical safety envelope and skill safety envelope, safety boundaries are defined from two dimensions: physical limits and subjective skill level.

[0065] In some embodiments, when the student's control instructions After being input into the system, the intent behind the instruction is understood through an intent prediction network based on LSTM (Linguistic Memory Time Metric). Then, combined with the position of the instruction relative to the double safety envelope, a tiered intervention strategy is executed. Specifically, the intent prediction network is a time-series classification model. Its input is a fixed time window. Continuous rod quantity sequence within The output is the probability distribution of the manipulation intention category C. The core of the LSTM network is the memory unit, whose state updates are determined by a gating mechanism; the hidden state at the last time step... The input is passed to a softmax classification layer to obtain the probability of the intent category:

[0066] ;

[0067] In the formula, It is the weight matrix of the output layer. It is the bias vector of the output layer.

[0068] Furthermore, the category of manipulation intent is obtained. In this example, it is mainly divided into three categories: regular training intent, high-mobility intent, and suspected misoperation intent.

[0069] The intervention steps involve implementing a tiered intervention strategy based on the position of the control command relative to the physical safety envelope, the skill safety envelope, and the control intent category. Specifically, the tiered intervention strategy includes: performing response passivation when the control command exceeds the skill safety envelope and the control intent category is a regular training intent; performing proactive correction intervention when the control command exceeds the skill safety envelope and the control intent category is a suspected misoperation intent; and performing safety locking when the control command exceeds the physical safety envelope.

[0070] Then, the flight response data output by the dedicated simulation model is monitored to see if it exceeds the preset safety envelope threshold. If it does, the shadow mode is activated, and the real-time control command is weighted and fused with the preset standard safety trajectory command to generate a corrective control quantity to replace the real-time control command.

[0071] In this embodiment, the intervention strategy is a rule-based, state-driven logic function, whose input is... The location relative to the boundaries of the physical safety envelope K and the skill safety envelope S, and the intent category. ,include:

[0072] The response is passively processed by linearly increasing the damping coefficient according to the degree to which the control command exceeds the safety envelope. This causes the angular acceleration response amplitude of the flight simulator to decrease proportionally, with the decrease ratio being positively correlated with the degree of exceedance. Specifically, when... and This is triggered on time. The core of this process is to increase the system's damping coefficient, thereby suppressing the aircraft's response. Let the original system's damping ratio be... The degree exceeding the skill safety envelope is defined as... The new damping ratio Increase proportionally:

[0073]

[0074] in It is a proportionality constant. In the dynamic model, this manifests as a decrease in the amplitude of the angular acceleration response. For example, for a roll channel, its angular acceleration... Compared with the original instruction The relationship has been modified to:

[0075] ;

[0076] in, For natural frequency, It is the angular velocity about the x-axis in the body coordinate system. The control command for the roll channel comes from the roll amount on the pilot's joystick (such as the normalized value of the remote control channel). This represents the moment of inertia of the drone about its x-axis. This increases the damping coefficient, dulls the angular acceleration response, and softens operations that exceed skill limits.

[0077] Active correction intervention includes vector compression of the control commands towards the skill safety envelope boundary, with the compressed commands replacing the original commands and inputting them into the dedicated simulation model. Specifically, when and When triggered, this process will convert the original control commands. Perform vector compression towards the boundary of the skill safety envelope. Let the distance from the envelope boundary be... The nearest point is Compressed instructions for:

[0078] ;

[0079] Here, λ is the compression factor, which can be a fixed value (such as 0.5) or dynamically adjusted according to the degree of excess. The compressed instruction replaces the original instruction and is input into the dedicated simulation model, thereby retaining some of the operational intent while bringing it back to the skill safety zone.

[0080] A safety lock smoothly switches control of the dedicated simulation model to an automatic hover state. Specifically, when... This is triggered on time, representing the highest level of intervention, smoothly switching control of the simulation model to an automatic hover state. The switching process uses a smooth transition function, for example, the final instruction. for:

[0081] ;

[0082] in, Smoothly growing from 0 to 1, It is the command required to maintain hovering.

[0083] In some embodiments, the graded intervention strategy further includes: monitoring whether the flight response data output by the dedicated simulation model exceeds a preset safety envelope threshold; if it does, activating the shadow mode, weighted and fused with the real-time control command and the preset standard safety trajectory command to generate a corrective control quantity to replace the real-time control command.

[0084] In this embodiment, the shadow mode is a safety mechanism that runs parallel to the aforementioned intervention strategy, namely, continuously monitoring the flight response data output by a dedicated simulation model. If the flight response data exceeds a preset safety envelope threshold, a shadow mode is activated. In shadow mode, the weighting coefficients of the weighted fusion are dynamically adjusted based on the degree to which the flight response data exceeds the safety envelope threshold. The greater the degree of exceedance, the higher the weight of the standard safety trajectory command. Its core is a weighted fusion algorithm that fuses the student's real-time control commands with preset standard safety trajectory commands. Let the standard safety trajectory command be... The fused correction control quantity for:

[0085] ;

[0086] Weighting coefficient Based on the extent to which flight response data exceeds the safety envelope threshold Dynamic adjustment The greater the degree of excess, The larger the value, the higher the weight of the standard safety trajectory command. Ultimately, It replaces real-time control commands for updating simulation models, ensuring the aircraft always operates within a safer envelope.

[0087] The teaching optimization steps involve collecting the control commands and flight response data, identifying error types based on an error feature library, generating teaching prompts, and decomposing the flight mission into basic control units. Features of each unit are extracted and input into a defect classifier to output defect type labels. Personalized compensation training tasks are then dynamically generated from a compensation training library based on these defect type labels. The error feature library includes at least attitude confusion features, throttle loss of control features, and beyond-visual-range panic features. The error type identification step specifically includes extracting the stick change rate of the control commands and the attitude lag angle from the flight response data, inputting the extracted features into a pre-trained classifier to output the corresponding error type.

[0088] Specifically, this step is the final stage in achieving the teaching objectives of the entire technical solution. It transforms flight data (control commands and flight responses) into structured teaching feedback and training tasks. More specifically, the defect classifier is a support vector machine or a multilayer perceptron, and the features include time-domain features, frequency-domain features, and phase features. The time-domain features include rise time, overshoot, and steady-state error; the frequency-domain features include the energy distribution of the stick power spectral density within a preset frequency band; and the phase feature is the peak delay time of the cross-correlation function between the input stick quantity and the output attitude angle. The compensation training library contains multiple basic compensation training units, each associated with a corresponding defect type label and training effect weight. The dynamic combination to generate personalized compensation training tasks includes selecting candidate units based on defect type labels and performing combination optimization with the goal of maximizing the overall training effect.

[0089] In some embodiments, the system collects control commands. With flight response data And it identifies errors based on a feature library of operational mistakes. First, feature vectors that characterize operational errors are extracted from the raw data. For example, including the rate of change of lever volume. Attitude hysteresis angle in flight response data etc., where the attitude lag angle can be calculated by the cross-correlation function between the input lever sequence and the output attitude angle sequence. get:

[0090] ;

[0091] In the formula, Indicates the roll angle of the drone. This indicates the attitude angle signal Shift forward on the timeline The result afterwards.

[0092] Therefore, the lag angle The time delay corresponding to the point where the cross-correlation function reaches its maximum value. These features are combined into a feature vector. Then, the data is fed into a pre-trained classifier (such as a support vector machine or a multilayer perceptron) to output the corresponding error type. (e.g., pose confusion, throttle loss, etc.). The decision function of the classifier can be expressed as:

[0093] ;

[0094] in, Category labels indicating the type of error, Represents conditional probability; It refers to the feature vector extracted from control commands and flight response data. Given the given conditions, this is the probability that the current operation belongs to error type c. This probability value is output by a pre-trained classifier (such as a support vector machine or a multilayer perceptron), and is usually normalized by a softmax function so that the sum of the probabilities of all categories is 1.

[0095] Furthermore, after identifying the error type, the system needs to generate targeted training tasks. First, the flight task (e.g., "figure-eight flight") is broken down into multiple basic control units (e.g., "level flight," "left turn," "right turn," "climb," etc.). For each unit, its multidimensional features are extracted. This includes: time-domain characteristics, such as rise time, overshoot, and steady-state error; frequency-domain characteristics, such as the energy distribution of the lever power spectral density within a preset frequency band; and phase characteristics, such as the peak delay time of the cross-correlation function between the input lever and the output attitude angle. These multi-dimensional characteristics... The input is fed into the defect classifier; in this embodiment, the defect classifier is a multi-classification model (such as a support vector machine or a multilayer perceptron), and its output is a defect type label. Based on the classifier learning, a model is constructed from the feature space to the defect label. The mapping.

[0096] Finally, the system is based on defect type labels. Personalized compensation training tasks are dynamically generated by combining elements from a compensation training library. The compensation training library contains multiple basic compensation training units, each unit i associated with a defect type label. and training effect weights The process of generating tasks can be viewed as a combinatorial optimization problem. Let the choice variable be... , indicating whether to select the i-th training unit. The goal is to select a set of units that maximizes the overall training effect while satisfying certain constraints, such as the total task duration. In this embodiment, the objective function can be defined as:

[0097] ;

[0098] The optimization goal is Set constraints ;

[0099] in, Let i be the time consumed for the i-th training unit. To the maximum allowed training duration, It is a decision variable, representing whether the i-th basic compensation training unit is selected for inclusion in the final generated personalized compensation training task. These are the training effect weights, representing the preset effectiveness coefficients of the i-th basic compensation training unit. It is the defect label of the training unit, indicating the defect type that the i-th basic compensation training unit specifically targets.

[0100] in This is a similarity function that measures the degree of match between the defects targeted by the training units and the current student's defects. The goal is to select a set of training units from the compensation training library that maximizes the sum of the effective training values ​​of these units for the student's defects. Solving this combinatorial optimization problem (e.g., using a greedy algorithm or dynamic programming) yields an optimal set of... This generates a personalized remedial training task, composed of multiple training units organically combined, specifically addressing the student's weaknesses. This task is then pushed to the student for the next stage of training, thus achieving precision and a closed-loop teaching process.

[0101] According to an embodiment of this application, an intelligent control system for drone simulation flight for university teaching includes a processor and a memory. The memory stores a computer program for implementing the intelligent control method for drone simulation flight for university teaching. The processor performs read and write operations on the memory.

[0102] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system.

[0103] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0104] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0105] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for intelligent control of unmanned aerial vehicle (UAV) simulated flight for university teaching, characterized in that: include: The model building steps include: obtaining the type information of the current teaching scenario; retrieving the corresponding environmental feature parameters based on the type information; resetting the parameters of the preset UAV dynamics model to generate a scenario-adaptive model; and collecting the student's control data during flight, identifying the student's personalized control feature parameters through an online parameter identification algorithm, embedding the personalized control feature parameters into the scenario-adaptive model to generate a unique simulation model. The safety construction steps include constructing a physical safety envelope and a skill safety envelope. The physical safety envelope is based on the physical limit parameters of the UAV, and the skill safety envelope is constructed based on the statistical distribution of students' historical normal operation data. The intention recognition step involves collecting the student's control command sequence in real time, inputting the control command sequence into an intention prediction network based on a long short-term memory network, and outputting the control intention category, which includes regular training intention, high mobility intention, and suspected misoperation intention. The intervention steps involve implementing a tiered intervention strategy based on the position of the control command relative to the physical safety envelope, the skill safety envelope, and the control intent category. The teaching optimization steps include collecting the control commands and flight response data, identifying error types based on the error operation feature database, and generating teaching prompt information. The flight mission is decomposed into basic control units, the features of each unit are extracted and input into a defect classifier to output defect type labels, and personalized compensation training missions are dynamically generated from the compensation training library based on the defect type labels.

2. The intelligent control method for unmanned aerial vehicle (UAV) simulated flight for university teaching according to claim 1, characterized in that, The tiered intervention strategy includes: When the control command exceeds the skill safety envelope and the control intention category is a regular training intention, a response passivation process is performed; when the control command exceeds the skill safety envelope and the control intention category is a suspected misoperation intention, an active correction intervention is performed; when the control command exceeds the physical safety envelope, a safety lock is performed. Then, the flight response data output by the dedicated simulation model is monitored to see if it exceeds the preset safety envelope threshold. If it does, the shadow mode is activated, and the real-time control command is weighted and fused with the preset standard safety trajectory command to generate a corrective control quantity to replace the real-time control command.

3. The intelligent control method for unmanned aerial vehicle (UAV) simulated flight for university teaching according to claim 2, characterized in that, The response passivation process includes: linearly increasing the damping coefficient according to the degree to which the control command exceeds the skill safety envelope, so that the angular acceleration response amplitude of the flight simulator is reduced proportionally, and the reduction ratio is positively correlated with the degree of exceedance; The active correction intervention includes: vector compression of the control command towards the skill safety envelope boundary, and inputting the compressed command into the dedicated simulation model instead of the original command; The security lock includes: smoothly switching control of the dedicated simulation model to an automatic hovering state.

4. The intelligent control method for unmanned aerial vehicle (UAV) simulated flight for university teaching according to claim 1, characterized in that, The environmental characteristic parameters include: A wind disturbance matrix is ​​introduced to address the playground environment; For airspace-restricted environments, a virtual potential field algorithm is superimposed on the UAV dynamics model to generate a reverse damping force when the aircraft approaches a preset electronic fence. The dynamic equations of the position loop are updated as follows: ； Where m is the mass and P is the position vector. To simulate the acceleration of an aircraft, This is the acceleration vector caused by wind disturbance. The reverse damping force is T, and the total thrust of the propeller is T. Let t be the rotation matrix from the body coordinate system to the ground coordinate system, and t be the time.

5. The intelligent control method for unmanned aerial vehicle (UAV) simulated flight for university teaching according to claim 1, characterized in that, The online parameter identification algorithm uses a recursive least squares algorithm to continuously update the personalized control feature parameters in a sliding time window manner. The width of the sliding time window is adaptively adjusted according to the student's training frequency.

6. The intelligent control method for unmanned aerial vehicle (UAV) simulated flight for university teaching according to claim 1 or 2, characterized in that, The skill safety envelope is constructed based on the statistical distribution of students' historical normal operation data, and includes: a Gaussian process regression model with control commands as input and the probability of reasonable operation as output. The boundary of the skill safety envelope corresponds to a preset reasonableness probability threshold. : In the formula, These are new control commands. This indicates that the Gaussian process model corresponds to the control commands. The predicted mean, It is a preset probability threshold for reasonableness; Furthermore, the control intention category is obtained through an intention prediction network, which is a time series classification model based on a long short-term memory network. Its input is a continuous lever sequence within a fixed time window, and its output is a probability distribution of regular training intention, high-maneuver intention, or suspected misoperation intention. The intent prediction network is a time series classification model built on an LSTM network, and its input is a fixed time window. Continuous rod quantity sequence within The output is the probability distribution of the manipulation intention category C: Among them, the probability of the intent category : ； In the formula, It is the weight matrix of the output layer. It is the bias vector of the output layer. It is the hidden state of the last time step; Manipulation intent category is .

7. The intelligent control method for unmanned aerial vehicle (UAV) simulated flight for university teaching according to claim 2, characterized in that, In the shadow mode, the weighting coefficients of the weighted fusion are dynamically adjusted according to the degree to which the flight response data exceeds the safety envelope threshold. The greater the degree of exceedance, the higher the weight of the standard safety trajectory command, including: Let the standard safety trajectory command be Weighted fusion corrected control quantity for: ； Among them, the weighting coefficient The system dynamically adjusts based on the degree to which flight response data exceeds the safety envelope threshold.

8. The intelligent control method for unmanned aerial vehicle (UAV) simulated flight for university teaching according to claim 2, characterized in that, The error feature library includes attitude confusion features, throttle loss features, and visual line-of-sight panic features; the steps for identifying error types specifically include: extracting the stick change rate of the control command and the attitude lag angle in the flight response data, and inputting the extracted features into a pre-trained classifier to output the corresponding error type.

9. The intelligent control method for unmanned aerial vehicle (UAV) simulated flight for university teaching according to claim 1, characterized in that, The defect classifier is a support vector machine or a multilayer perceptron, and the features include time-domain features, frequency-domain features, and phase features. The time-domain features include rise time, overshoot and steady-state error; the frequency-domain features include the energy distribution of the lever power spectral density in a preset frequency band; and the phase feature is the peak delay time of the cross-correlation function between the input lever and the output attitude angle. The compensation training library contains multiple basic compensation training units, each associated with a corresponding defect type label and training effect weight; the dynamic combination to generate personalized compensation training tasks includes filtering candidate units based on defect type labels and performing combination optimization with the goal of maximizing the overall training effect. The objective function is defined as: ； The optimization goal is Set constraints ; in, Let i be the time consumed for the i-th training unit. To the maximum allowed training duration, It is a decision variable. These are training effect weights. These are the defect labels of the training units. It is a similarity function. Let be the time taken for the i-th training unit.

10. An intelligent control system for unmanned aerial vehicle (UAV) flight simulation for university teaching, characterized in that: The device includes a processor and a memory, wherein the memory stores a computer program for implementing the intelligent control method for simulated flight of unmanned aerial vehicles for college teaching as described in any one of claims 1-9, and the processor performs read and write operations on the memory.

Images