Flight trajectory control method and system based on operation state sensing and instruction learning

Abstract

The invention belongs to the technical field of data processing, relates to a flight path control method and system based on operation state sensing and instruction learning, and aims to solve the problems of difficulty in deviation correction under complex disturbance, weak dynamic adaptability and insufficient fusion of instruction and state sensing in a traditional method. The method comprises the following steps: acquiring flight data; constructing a joint input tensor, and obtaining a hidden state sequence through an encoder; operation probability distribution is obtained based on the prediction model, and then a predicted roll angle instruction is determined; a basic track is obtained based on the roll angle instruction, disturbance compensation amount is obtained based on the wind disturbance sequence, and a lateral offset track is obtained through residual error correction network correction; under the constraint of the safety envelope, the roll angle instruction correction is determined, and the transverse motion state is corrected; and performing security boundary detection after correction. Through real-time sensing of the operation state and fusion instruction learning, disturbance and characteristic changes can be accurately captured, and the control efficiency, adaptability and intelligent level are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of flight control, and in particular relates to a flight trajectory control method and system based on operation state perception and instruction learning. Background Art

[0002] In the aviation safety system, and particularly in the aerospace sector, precise control of flight trajectory is one of the core technologies for ensuring safe and efficient aircraft operations. As mission complexity increases, aircraft face a variety of dynamic interference factors during actual flight. These include environmental disturbances such as atmospheric turbulence and gusts, as well as inherent variations in the aircraft's own characteristics, such as aging of the powertrain and actuator delays. The combined effect of these factors can cause deviations between the actual flight trajectory and the pre-set instructions. Failure to promptly correct these deviations can lead to excessive trajectory deviations, increased energy consumption, and even accidents. Traditional flight trajectory control methods often rely on pre-set mathematical models, implementing trajectory planning and correction by establishing fixed-parameter control equations. However, these methods place extremely high demands on model accuracy. When faced with complex and changing operational conditions (such as sudden command adjustments or extreme weather conditions), fixed models struggle to adaptively capture dynamic changes, making control lag or overshoot a common problem. Furthermore, traditional methods have limited learning capabilities for operator input commands and are unable to effectively correlate the mapping between historical commands and trajectory responses, resulting in insufficient control robustness during multi-task switching or under atypical operating conditions. In recent years, with the development of intelligent perception and machine learning technologies, data-driven control methods have gradually become a research hotspot. However, most existing solutions focus solely on state perception or command optimization, failing to achieve a deep integration of dynamic perception of operating states and command learning. This makes it difficult to balance real-time and accurate trajectory control.

[0003] Therefore, how to construct a technical solution that can accurately perceive the flight operation status and achieve trajectory stability control under dynamic interference has become a key issue that needs to be urgently solved in this field. Summary of the Invention

[0004] In order to solve the above-mentioned problems in the prior art, namely, the difficulty in correcting trajectory deviations under complex disturbances in flight control methods, weak model dynamic adaptability, and insufficient fusion of instructions and state perception, the first aspect of the present invention proposes a flight trajectory control method based on operation state perception and instruction learning, the method comprising the following steps: Acquiring flight data, the flight data including a wind disturbance sequence, an operation sequence, and a reference roll angle instruction sequence; A joint input tensor is constructed based on the operation sequence and the baseline roll angle command sequence, and a hidden state sequence is obtained after being processed by the encoder. The hidden state sequence is input into a preset prediction model to obtain the operation probability distribution, and then the predicted roll angle command is determined. A kinematic model is constructed based on the roll angle command to obtain a basic trajectory, a wind disturbance model is constructed based on the wind disturbance sequence to obtain a disturbance compensation amount, and a composite result of the basic trajectory and the disturbance compensation amount is corrected by a residual correction network to obtain a lateral offset trajectory; Adjusting the range of the safety envelope according to the lateral offset trajectory, and determining the roll angle command correction value based on the lateral offset trajectory under the constraints of the safety envelope, and correcting the lateral motion state of the vehicle according to the roll angle command correction value; After the correction, a safety boundary check is performed. If the test result does not fall within the safety envelope, the correction amount is recalculated and the correction operation is repeated. If it meets the safety envelope requirements, continuous monitoring is performed.

[0005] In some preferred embodiments, the method for constructing a joint input tensor based on the operation sequence and the reference roll angle instruction sequence and obtaining a hidden state sequence after being processed by the encoder is as follows: The operation sequence and the reference roll angle command sequence are processed using a learnable embedding function, mapped into the same semantic space, and the two obtained embedding vectors are concatenated in dimension to form a joint input tensor; The joint input tensor is used as an input sequence, and after being processed by an encoder for extracting temporal features, a hidden state sequence is obtained, where the hidden state sequence contains context information and time dependency of the input sequence.

[0006] In some preferred embodiments, the predicted roll angle command is determined by: Based on the hidden state sequence, obtaining an operation probability distribution through a prediction model; The operation type with the highest probability under the operation probability distribution is used as the predicted editing operation type of the current time step; Based on the predicted editing operation type and in combination with a preset operation set, a roll angle instruction is determined.

[0007] In some preferred embodiments, based on the predicted editing operation type and in combination with a preset operation set, a roll angle instruction is determined by: Construct an operation set based on historical data: ; in, is the correction amount obtained by clustering historical data. k Roll angle history correction, 、 、 、 Represents different operation types. Indicates that the base roll angle command remains unchanged. Indicates adding a k Class correction amount, Indicates that the kth type of correction is reduced based on the baseline instruction. Indicates direct replacement of the base instruction; k =1....K, K is the number of clusters; Determine the roll angle instruction based on the correspondence between the measurement editing operation type and the operation set : ; ; in, is the reference roll angle command, For offset compensation of REPLACE operation type, is hidden state, To map high-dimensional hidden features into linear weights of roll angle commands, To predict the type of editing operation, is the probability distribution, To find the editing operation type with the highest probability under the operation probability distribution.

[0008] In some preferred embodiments, the method for obtaining the lateral offset trajectory is as follows: The kinematic model is used as input to calculate the basic trajectory of the lateral deviation of the carrier in the undisturbed state through the preset kinematic equations. ; ; The wind disturbance model is based on the temporal characteristics of the wind disturbance sequence and outputs the disturbance compensation for the lateral displacement of the carrier by fitting the wind field action law. ; ; The residual correction network obtains the residual correction by learning the model error law in the historical trajectory data ; ; ; Based on the basic trajectory, disturbance compensation, and residual correction, the corrected lateral offset trajectory is obtained: ; in, is the airspeed constant, is the time step, is the wind direction angle, is the cosine of the wind direction angle, is the predicted roll angle command of the current time step after sinusoidalization, is the predicted lateral offset of the previous time step, is the predicted roll angle command for the current time step, is the instantaneous wind speed and volume of the wind disturbance at the current time step, is the encoder output hidden layer dimension, is the input feature vector of the residual correction network, MLP is a multi-layer perceptron, are the learnable parameters in MLP, is hidden state, is the set of real numbers.

[0009] In some preferred embodiments, the flight data further includes the current lateral deviation trajectory of the vehicle, flight speed, deviation safety threshold, and envelope boundary slope; and the roll angle command correction amount is determined based on the lateral deviation trajectory under the safety envelope constraint, and the method is as follows: Based on the lateral offset trajectory, the roll angle command correction sequence is obtained by reverse calculation; Based on the genetic algorithm, the roll angle command correction sequence is encoded as an individual of the genetic algorithm. The population iteration is combined with the fitness function to screen out the roll angle command correction sequence corresponding to the best individual as the candidate optimal solution. Based on the Lagrange multiplier method, with the minimization of the pilot's operation amount as the objective function and the safety envelope as the constraint condition, the optimal correction amount among the candidate optimal solutions is solved as the roll angle command correction amount. A second aspect of the present invention provides a flight trajectory control system based on operation state perception and instruction learning, the system comprising: a data acquisition module configured to collect flight data in real time, the flight data including a current lateral deviation trajectory, flight speed, deviation safety threshold, envelope boundary slope, wind disturbance sequence, operation sequence, and reference roll angle command sequence; a preprocessing module configured to construct a joint input tensor based on the operation sequence and the reference roll angle command sequence, and obtain a hidden state sequence after being processed by the encoder; a prediction module configured to input the hidden state sequence into a preset prediction model to obtain an operation probability distribution, and then determine a roll angle command based on the operation probability distribution; an offset analysis module configured to construct a kinematic model based on the roll angle command to obtain a basic trajectory, construct a wind disturbance model based on the wind disturbance sequence to obtain a disturbance compensation amount, and correct a composite result of the basic trajectory and the disturbance compensation amount through a residual correction network to obtain a lateral offset trajectory; A safety envelope adjustment module is configured to analyze a deviation trend according to a lateral deviation trajectory and adjust a range of a safety envelope; a correction module configured to determine a roll angle command correction value based on the lateral offset trajectory under the constraints of the safety envelope, and correct the lateral motion state of the carrier according to the roll angle command correction value; The feedback module is configured to perform safety boundary detection after correction. If the detection result does not fall within the safety envelope, it returns to recalculate the correction amount and repeat the correction operation. If it meets the safety envelope requirements, it continues to monitor.

[0010] Beneficial effects of the present invention: 1. The pilot's stick push operation and the reference roll angle command are embedded and fused to form a joint input tensor, which is then encoded using an encoder. A learning mechanism is used to continuously optimize command parsing and trajectory correction strategies, reducing reliance on manually preset parameters and achieving a deep integration of dynamic perception of operational status and command learning. 2. A command editing learning mechanism was built, defining an operation set containing multiple editing operations. By predicting the operation type at each time step, the baseline roll angle command can be precisely edited and adjusted, achieving optimized command generation and improving the flexibility and adaptability of flight control. 3. Combining physical kinematic equations and residual correction networks, while also introducing wind disturbance modeling, allows trajectory prediction to reflect physical laws while adapting to environmental changes, improving anti-interference capabilities and ensuring safe flight in complex weather conditions. 4. A dynamically weighted joint training strategy is proposed. By dynamically adjusting the loss weights of tasks such as editing operations, trajectory prediction, and wind disturbance processing, this strategy addresses the problem of unbalanced losses among tasks in multi-task learning, improves the overall performance of the model, and achieves effective balance and collaborative optimization among different tasks. 5. Genetic algorithms and Lagrange multiplier method are used to optimize the roll angle command correction under the safety envelope constraints, ensuring that the aircraft's lateral deviation trajectory is accurately controlled within a safe range, achieving real-time protection and dynamic adjustment of flight safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Other features, objects and advantages of the present application will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings: Figure 1 This is a flow chart of a flight trajectory control method based on operation state perception and instruction learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] The present application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the relevant invention are shown in the accompanying drawings.

[0013] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0014] The present invention provides a flight trajectory control method based on operating state perception and command learning. By perceiving the operating state in real time and integrating command learning, this method can accurately capture disturbances and characteristic changes to reduce trajectory deviation and improve dynamic response speed. It breaks through the limitations of fixed models and adaptively responds to complex working conditions to enhance robustness. While improving control efficiency, adaptability and intelligence level, it provides support for the efficient and safe execution of flight missions.

[0015] A flight trajectory control method based on operation state perception and instruction learning of the present invention comprises the following steps: Acquiring flight data, the flight data including a wind disturbance sequence, an operation sequence, and a reference roll angle instruction sequence; A joint input tensor is constructed based on the operation sequence and the baseline roll angle command sequence, and a hidden state sequence is obtained after being processed by the encoder. The hidden state sequence is input into a preset prediction model to obtain the operation probability distribution, and then the predicted roll angle command is determined. A kinematic model is constructed based on the roll angle command to obtain a basic trajectory, a wind disturbance model is constructed based on the wind disturbance sequence to obtain a disturbance compensation amount, and a composite result of the basic trajectory and the disturbance compensation amount is corrected by a residual correction network to obtain a lateral offset trajectory; Adjusting the range of the safety envelope according to the lateral offset trajectory, and determining the roll angle command correction value based on the lateral offset trajectory under the constraints of the safety envelope, and correcting the lateral motion state of the vehicle according to the roll angle command correction value; After the correction, a safety boundary check is performed. If the test result does not fall within the safety envelope, the correction amount is recalculated and the correction operation is repeated. If it meets the safety envelope requirements, continuous monitoring is performed.

[0016] In order to more clearly explain the flight trajectory control method based on operation state perception and instruction learning of the present invention, the following is combined with Figure 1 Each step in the embodiment of the present invention is described in detail.

[0017] The flight trajectory control method based on operation state perception and instruction learning of the first embodiment of the present invention includes steps S1 to S5, each of which is described in detail as follows: S1. Acquire flight data, including wind disturbance sequence , operation sequence and the reference roll angle command sequence ; in, The variable of the reference roll angle instruction sequence, which represents the time step t The roll angle command generated according to the basic flight control logic is the pilot's initial expected value of the aircraft's roll angle. The dimension is T, that is, the reference roll angle command value corresponding to T time steps.

[0018] Preferably, the flight data also includes the lateral deviation trajectory, flight speed, deviation safety threshold, and envelope boundary slope at the current moment.

[0019] In this embodiment, the operation sequence is obtained through a flight parameter recorder and data gloves worn by the pilot; the wind disturbance sequence is obtained by recording weather radar and pitot tube, and only the lateral (perpendicular to the runway) component is considered; and the carrier is an aircraft.

[0020] In this embodiment, the reference roll angle instruction sequence is obtained by the PD controller: ; in, is the current lateral offset error, is the rate of change of the current lateral offset error, are the proportional term parameters and differential term parameters of the controller respectively.

[0021] This application innovatively embeds and fuses the pilot's stick operation with the reference roll angle command to form a joint input tensor, which is encoded using the Transformer architecture to achieve dynamic perception and modeling of the pilot's operational intentions, providing a basis for subsequent command editing.

[0022] S2. Construct a joint input tensor based on the operation sequence and the reference roll angle command sequence, and obtain a hidden state sequence after processing by the encoder; input the hidden state sequence into the preset prediction model to obtain the operation probability distribution, and then determine the predicted roll angle command .

[0023] Preferably, the method of constructing a joint input tensor based on the operation sequence and the reference roll angle instruction sequence and obtaining a hidden state sequence after being processed by an encoder is as follows: The operation sequence and the reference roll angle instruction sequence are processed using a learnable embedding function, mapped to the same semantic space, and the two embedded vectors obtained are concatenated in dimension to form a joint input tensor. ;

[0024] in, is a learnable embedding function, dThe embedding dimension is the vector dimension after the original pilot's stick operation and reference roll angle command information are converted through the embedding function. It is used to map different types of input data into the same semantic space. T ×2 d represents the dimensions of the joint input tensor, 2d The dimension of the concatenated vector; The joint input tensor is used as the input sequence and processed by the Transformer encoder for extracting temporal features to obtain the hidden state sequence , the hidden state sequence contains the context information and time dependency of the input sequence.

[0025] Further preferably, the position coding adopts relative time coding, and the method is as follows: ; ;

[0026] in, is a hidden state sequence containing the contextual information and temporal dependencies of the input sequence, with a dimension of ; is the hidden state sequence at time step t The corresponding hidden state is an element in the encoder output sequence; d h It is the output hidden layer dimension of the Transformer encoder. The position encoding values ​​at different time steps and dimensions are calculated through sine and cosine functions. It is used to provide time sequence information to the model and help the model understand the time sequence relationship in the input sequence.

[0027] Through relative time encoding, it plays a role when the model processes input sequences (such as the joint input tensor composed of the pilot operation sequence and the reference roll angle instruction sequence), ensuring that the model can grasp the logical sequence of input information in the time dimension and provide contextual support in the time dimension for subsequent tasks such as editing operation prediction.

[0028] Preferably, the predicted roll angle command is determined by: Construct an operation set based on historical data: ; in, The historical correction value of the roll angle is obtained by clustering the correction values ​​in the historical data. The operation set is obtained by enumerating the actions of the typical correction values. 、 、 、 Represents different types of operations, which are used to define the operations that the model can perform on the reference roll angle command at each time step. k =1....K; Based on the hidden state sequence, the predicted time step is obtained through the prediction model. t The probability distribution of the operation ; ; are the learnable parameters of the model, is the weight matrix, is the bias term, which is used to convert the hidden state of the encoder output Convert to a distribution of operation probabilities; Is the dimension identifier of the real vector space, the output probability vector have dimensions, each dimension corresponds to the probability of an operation type; The operation type with the highest probability under the operation probability distribution is used as the predicted editing operation type of the current time step ; ; Based on the predicted editing operation type, combined with the preset operation set, and according to the corresponding relationship between the predicted editing operation type and the operation set, the roll angle instruction is determined. : ; in, is the reference roll angle command, For offset compensation of REPLACE operation type, is hidden state, To map high-dimensional hidden features into linear weights of roll angle commands, To predict the type of editing operation, is a probability distribution.

[0029] In this embodiment, the historical data is clustered to obtain As the core, define the instruction editing actions that the model can execute at each time step: Indicates that the base roll angle command remains unchanged (no correction), Indicates adding the kth type of correction amount based on the baseline instruction. Indicates that the kth type of correction is reduced based on the baseline instruction. Indicates direct replacement of the baseline instruction (applicable to scenarios requiring significant adjustments).

[0030] Further preferably, domain knowledge constraints are applied to the predicted editing operation type, and the probability distribution is renormalized: .

[0031] in, t is the time step, T is the total number of time steps, and the operation distribution of the last five time steps of the sequence is set to 0. An operation set containing multiple editing operations is defined. By predicting the operation type of each time step, the reference roll angle command is accurately edited and adjusted to achieve optimized command generation, thereby improving the flexibility and adaptability of flight control.

[0032] S3. Construct a kinematic model based on the roll angle command to obtain a basic trajectory, construct a wind disturbance model based on the wind disturbance sequence to obtain a disturbance compensation amount, and correct the synthesis result of the basic trajectory and the disturbance compensation amount through a residual correction network to obtain a lateral offset trajectory. .

[0033] The method for obtaining the lateral offset trajectory is as follows: The kinematic model uses the roll angle command as input and calculates the basic lateral deviation trajectory of the carrier in the undisturbed state through the preset kinematic equation. ; ; The wind disturbance model is based on the temporal characteristics of the wind disturbance sequence and outputs the disturbance compensation for the lateral displacement of the carrier by fitting the wind field action law. ; ; Define the input feature vector, which includes the predicted roll angle command at the current time step, the predicted lateral offset at the previous time step, the wind disturbance at the current time step, and the hidden state of the encoder output; ; The residual correction network obtains the residual correction by learning the model error law in the historical trajectory data ; ; Based on the basic trajectory, disturbance compensation, and residual correction, the corrected lateral offset trajectory is obtained: ; in, is the airspeed constant, is the time step, is the wind direction angle, is the cosine of the wind direction angle, is the predicted roll angle command of the current time step after sinusoidalization, is the predicted lateral offset of the previous time step, is the predicted roll angle command for the current time step, is the instantaneous wind speed and volume of the wind disturbance at the current time step, is the encoder output hidden layer dimension, is the input feature vector of the residual correction network, MLP is a multi-layer perceptron, is hidden state, is the set of real numbers.

[0034] By combining the physical kinematic equations and the residual correction network, and introducing wind disturbance modeling, the trajectory prediction can both reflect physical laws and adapt to environmental changes, thus improving anti-interference capabilities and ensuring the safe flight of aircraft in complex weather conditions.

[0035] Preferably, the loss function of the lateral offset trajectory is for: ; ; ; ; in, is the editing operation loss, is the trajectory loss, is the wind disturbance consistency regularization term; is the wind disturbance impact weight, is the weight coefficient, i= 1, 2, 3; For real editing operations, is the total number of time steps, is the probability distribution, For time t The lateral offset trajectory, For time t The true offset trajectory, For time t Ideal lateral deviation trajectory under no wind conditions.

[0036] Further preferably, the weight of the loss function adopts dynamic weighting: ; in, is the task uncertainty parameter, i=1, 2, 3 correspond to 、 、 , which is used to balance the proportion of editing operation loss, trajectory loss and wind disturbance consistency regularization term in the total loss function. It is a task uncertainty parameter that reflects the model's uncertainty estimation of different tasks (such as editing operation prediction, trajectory prediction and wind disturbance processing, etc.). It is used to dynamically adjust the weight of the corresponding loss in the total loss. When the model has a large uncertainty for a certain task, the corresponding weight will decrease, and vice versa.

[0037] By dynamically adjusting the loss weights for tasks such as editing, trajectory prediction, and wind disturbance handling, the problem of unbalanced losses across tasks in multi-task learning is addressed, the overall performance of the model is improved, and effective balance and coordinated optimization between different tasks is achieved. At the same time, by decoupling physical kinematics from the residual correction network and deeply integrating wind disturbance modeling, the "pure control error" is accurately extracted, allowing the model to focus on improving its own strategy during optimization rather than the interference of wind disturbances. This achieves a synergy between physical correctness and anti-disturbance flexibility, making trajectory predictions in complex weather conditions more accurate and stable.

[0038] S4. Adjust the range of the safety envelope according to the lateral offset trajectory, and determine a roll angle command correction based on the lateral offset trajectory under the constraints of the safety envelope, and correct the lateral motion state of the carrier according to the roll angle command correction.

[0039] Preferably, the range of the safety envelope is adjusted according to the lateral offset trajectory by: Different flight phases (such as takeoff, climb, cruise, descent, landing, etc.) have different sensitivities and safety requirements for lateral deviation. For example, during takeoff and landing, the aircraft has a lower tolerance for lateral deviation, so the safety control envelope will be stricter. Based on the initial heading deviation trajectory, the deviation trend and acceleration are determined, and the predicted lateral deviation trajectory is used to identify potential danger points. Based on the excursion trend, acceleration, wind speed and corresponding flight phase, a safety envelope is constructed in combination with the preset excursion safety threshold. The safety envelope is readjusted based on the predicted heading deviation trajectory obtained from the real-time data received. The predicted lateral deviation trajectory already includes the effect of wind on the aircraft's lateral motion. Therefore, the safety envelope can be adjusted according to environmental conditions, such as appropriately narrowing the safety envelope in strong crosswind conditions.

[0040] Further preferably, in this embodiment, the lateral deviation safety threshold is a key parameter for evaluating the pilot's safety control envelope, and is determined according to the runway width, flight mission type, etc. The safety threshold can be set to half the runway width ; Based on the predicted lateral deviation trajectory, find the time points that exceed the lateral deviation safety threshold , that is, satisfy These points are called potential danger points, which are used to quantify risks, trigger correction mechanisms, and actively improve safety indexes.

[0041] Further preferably, the velocity and acceleration of the lateral deviation trajectory are calculated to analyze the deviation trend, and the method is as follows: ; ; in, is the offset rate, is the deviation acceleration; if the lateral deviation speed and acceleration are large, it may mean that the aircraft's lateral deviation tendency is aggravated, and the safety envelope needs to be considered more carefully.

[0042] In this embodiment, a safety envelope is graphically represented, with time or spatial position as the horizontal axis and lateral offset as the vertical axis, plotting the upper and lower limits of the safety envelope. The area within this envelope is the safe zone, and the area outside this envelope is the dangerous zone.

[0043] Preferably, the roll angle command correction amount is determined based on the lateral offset trajectory under the safety envelope constraint, and the method is: The flight data also includes the current lateral deviation trajectory, flight speed, deviation safety threshold, and envelope boundary slope of the vehicle; Based on the lateral offset trajectory, the roll angle command correction sequence is obtained by reverse calculation , and initialize the parameters; Based on the genetic algorithm, the roll angle command correction sequence is encoded as an individual of the genetic algorithm. The population iteration is combined with the fitness function to screen out the roll angle command correction sequence corresponding to the best individual as the candidate optimal solution. Based on the Lagrange multiplier method, with the minimization of the pilot's operation amount as the objective function and the safety envelope as the constraint condition, the optimal correction amount among the candidate optimal solutions is solved as the roll angle command correction amount.

[0044] Preferably, based on the lateral offset trajectory, the roll angle instruction correction sequence is obtained by reverse deduction, and the method is as follows: The dynamic relationship between lateral offset and roll angle is established, the offset error is defined, the correction value is calculated using the PID controller, the discretization result is obtained to obtain the correction value sequence, and then the parameters are iteratively optimized.

[0045] Further preferably, the correction amount sequence is dynamically expanded: ; in, The maximum / minimum value of the difference in the instruction correction amount sequence is the step coefficient that evenly divides the "maximum-minimum difference interval" into N parts. When the usage frequency exceeds the threshold, the new correction amount is activated.

[0046] The construction of discrete adjustment values ​​through “extreme difference segmentation” is a preparation for the subsequent selection / generation of specific reference roll angle instructions, which is a key step in “defining candidate instruction values” in the instruction generation process. It becomes a discrete set containing "original deviation + evenly divided step size". Values ​​can be selected from this set to generate discrete reference roll adjustment instructions (for example, adjusting the instruction value according to a fixed step size, which meets the engineering requirement that "instructions must be discrete and controllable").

[0047] Further preferably, the parameter initialization includes setting initial parameters of the genetic algorithm, such as the population size (e.g., 100-200 individuals), maximum number of iterations (e.g., 50-100 generations), crossover probability (0.6-0.9), and mutation probability (0.01-0.1), based on the aircraft model and flight phase. Simultaneously, the initial iteration accuracy (e.g., 1e-6) and maximum number of iterations (e.g., 1000) of the Lagrange multiplier method are determined.

[0048] More preferably, the optimal solution candidate is screened based on a genetic algorithm, and the method is as follows: 1) Individual Encoding and Population Initialization: The roll angle command correction sequence is encoded as an individual in the genetic algorithm, with each individual representing a possible correction strategy. When initializing the population, multiple different correction sequences are randomly generated as initial individuals to ensure population diversity. 2) Fitness Function Calculation: For each individual, simulate the change in lateral deviation trajectory after the corresponding roll angle command correction is applied to the aircraft. Calculate the degree of deviation from the safety envelope, pilot operation parameters (such as the sum of the absolute values ​​of the corrections and the frequency of operation), and flight state stability (such as the rate of change of lateral deviation and the rate of change of roll angle command). These parameters are combined to construct a fitness function. 3) Selection operation: Based on individual fitness, individuals with higher fitness are selected as parents, so that they have a higher probability of passing genes to the next generation; 4) Crossover and mutation operations: Perform a crossover operation on the selected parent individuals, randomly exchanging some gene fragments to generate new individuals; at the same time, mutate the individual genes with a certain probability, randomly changing the values ​​of some correction variables to increase population diversity and exploration capabilities; 5) Iterative update and termination condition judgment: Repeat the selection, crossover, and mutation operations to continuously update the population. When the maximum number of iterations is reached or the fitness of individuals in the population converges (for example, the change in the optimal fitness of the population over multiple generations is less than a certain threshold), the iteration is terminated. The roll angle command correction value sequence corresponding to the optimal individual at this time is output as the candidate optimal solution.

[0049] Preferably, the genetic algorithm is optimized: The comparison results between the current lateral deviation trajectory and the safety envelope (whether it exceeds the limit, the duration of the exceedance / the amount of the offset) are used as constraints and encoded into the fitness function of the genetic algorithm (the out-of-bounds strategy is directly marked as invalid individuals and eliminated). The potential risk level of the subsequent short-term lateral deviation trend (such as the predicted rapid expansion of the offset under strong winds) is used as the optimization target to drive the genetic algorithm to search for a control strategy that "can suppress the spread of risk" (the higher the risk, the greater the fitness weight of the corresponding strategy).

[0050] As an option, in this embodiment, the fitness function is: ; in, is the weight coefficient, which is determined according to the flight mission and safety requirements.

[0051] As an option, in this embodiment, a method such as roulette wheel selection or tournament selection is used to select individuals with higher fitness as parents.

[0052] Evaluate the effectiveness and robustness of the search through the population evolution characteristics of the genetic algorithm: Population diversity: Monitor the differences between individuals (control strategies) during iteration (e.g., the range and frequency of roll angle corrections) to prevent the algorithm from prematurely converging to a local optimum (e.g., focusing only on staying within bounds, which can lead to wild fluctuations in control variables). Fitness convergence: Track the fitness changes of the optimal individual (combined with "safety envelope fit + risk suppression effect") to determine whether the algorithm is stably approaching the global optimum (continuous improvement in fitness with small fluctuations indicates that the search is effective).

[0053] More preferably, the optimal correction amount is solved based on the Lagrange multiplier method, which is as follows: 1) Objective function and constraint condition construction: The objective function is to minimize the pilot's operation (e.g., minimize the sum of the squares of the roll angle command corrections): ; At the same time, the corrected lateral deviation trajectory must be within the safety envelope, that is, for each time point t ,satisfy ,in To predict lateral drift, is the lateral offset change caused by the correction amount; Introducing Lagrange multipliers Construct the Lagrangian function: ; 2) Optimal solution: Derivatives of the Lagrangian functions are taken to derive a set of optimality condition equations. Numerical methods are used to solve these equations to obtain the optimal roll angle command correction sequence that satisfies the constraints, ensuring that the aircraft's lateral excursion is precisely controlled within the safety envelope. 3) Correction Validation and Application: The optimal correction is applied to the current roll angle command to simulate and verify whether the aircraft's lateral deviation trajectory remains stable within the safety envelope. If the requirements are met, the corrected roll angle command is sent to the flight control system for execution. If not, the constraints or optimization parameters are adjusted appropriately, and the solution and verification are repeated until a feasible solution is obtained.

[0054] The genetic algorithm provides global exploration capabilities, the Lagrange multiplier method ensures constraint satisfaction and accuracy, and can dynamically adjust the correction amount through real-time feedback. The genetic algorithm and Lagrange multiplier method are used to optimize the roll angle command correction amount under the safety envelope constraints, ensuring that the aircraft's lateral deviation trajectory is accurately controlled within a safe range. This solves the limitations of traditional single methods in dynamic constraint response, multi-constraint coordination and adaptation to complex working conditions, and achieves stable control of the trajectory within the safety envelope.

[0055] S5. After the correction, perform safety boundary detection. If the detection result does not fall within the safety envelope, return to recalculate the correction amount and repeat the correction operation. If it meets the safety envelope requirements, continue monitoring.

[0056] Continuously monitor the aircraft's actual lateral deviation trajectory, flight attitude and other status parameters to evaluate the control effect; if it is found that the actual trajectory still deviates from the expected or new interference factors appear, promptly feed the new data back to the genetic algorithm and Lagrange multiplier method modules, restart the optimization process, and dynamically adjust the roll angle command correction amount to ensure that the aircraft remains stable within the safety envelope.

[0057] Preferably, the correction effect is evaluated according to performance evaluation indicators, which include: Instruction generation efficiency: ; Trajectory safety deviation: ; Wind disturbance suppression rate: ; in, It is an ideal trajectory without wind; SDI>1 indicates out of bounds.

[0058] Although the various steps in the above embodiment are described in the above-mentioned order, those skilled in the art will understand that in order to achieve the effect of this embodiment, different steps do not have to be executed in such an order. They can be executed simultaneously (in parallel) or in a reverse order. These simple changes are within the scope of protection of the present invention.

[0059] A flight trajectory control system based on operation state perception and instruction learning according to a second embodiment of the present invention includes: a data acquisition module configured to collect flight data in real time, the flight data including a current lateral deviation trajectory, flight speed, deviation safety threshold, envelope boundary slope, wind disturbance sequence, operation sequence, and reference roll angle command sequence; a preprocessing module configured to construct a joint input tensor based on the operation sequence and the reference roll angle command sequence, and obtain a hidden state sequence after being processed by the encoder; a prediction module configured to input the hidden state sequence into a preset prediction model to obtain an operation probability distribution, and then determine a roll angle command based on the operation probability distribution; an offset analysis module configured to construct a kinematic model based on the roll angle command to obtain a basic trajectory, construct a wind disturbance model based on the wind disturbance sequence to obtain a disturbance compensation amount, and correct a composite result of the basic trajectory and the disturbance compensation amount through a residual correction network to obtain a lateral offset trajectory; A safety envelope adjustment module is configured to analyze a deviation trend according to a lateral deviation trajectory and adjust a range of a safety envelope; a correction module configured to determine a roll angle command correction value based on the lateral offset trajectory under the constraints of the safety envelope, and correct the lateral motion state of the carrier according to the roll angle command correction value; The feedback module is configured to perform safety boundary detection after correction. If the detection result does not fall within the safety envelope, it returns to recalculate the correction amount and repeat the correction operation. If it meets the safety envelope requirements, it continues to monitor.

[0060] It should be noted that the flight trajectory control system based on operational state perception and command learning provided in the above embodiment is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be further decomposed or combined. For example, the modules in the above embodiment can be combined into one module or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are merely for the purpose of distinguishing the modules or steps and are not to be regarded as improper limitations on the present invention.

[0061] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working process and related instructions of the system described above can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.

[0062] An electronic device according to a third embodiment of the present invention includes: at least one processor; and a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the above-mentioned flight trajectory control method based on operation state perception and instruction learning.

[0063] A computer-readable storage medium according to a fourth embodiment of the present invention stores computer instructions, which are used to be executed by a computer to implement the above-mentioned flight trajectory control method based on operation state perception and instruction learning.

[0064] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes and related instructions of the electronic device and computer-readable storage medium described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0065] Those skilled in the art should be able to appreciate that the modules and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two, and the programs corresponding to the software modules and method steps can be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art. In order to clearly illustrate the interchangeability of electronic hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0066] Computer program code for performing the operations of the present application may be written in one or more programming languages, or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0067] The flow charts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, program segment or a part of code, and the module, program segment or a part of code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented by a dedicated hardware-based system that performs the specified function or operation, or can be implemented by a combination of dedicated hardware and computer instructions.

[0068] The terms "first", "second", etc. are used to distinguish similar objects, rather than to describe or indicate a particular order or sequence.

[0069] The term "comprise" or any other similar term is intended to cover non-exclusive inclusion such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0070] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

Claims

1. A flight trajectory control method based on operation state perception and instruction learning, characterized in that: The method comprises the following steps: Acquiring flight data, the flight data including a wind disturbance sequence, an operation sequence, and a reference roll angle instruction sequence; A joint input tensor is constructed based on the operation sequence and the reference roll angle command sequence. After being processed by the encoder, a hidden state sequence is obtained. The hidden state sequence is input into a preset prediction model to obtain the operation probability distribution, and then the predicted roll angle command is determined. A kinematic model is constructed based on the roll angle command to obtain a basic trajectory, a wind disturbance model is constructed based on the wind disturbance sequence to obtain a disturbance compensation amount, and a composite result of the basic trajectory and the disturbance compensation amount is corrected by a residual correction network to obtain a lateral offset trajectory; Adjusting the range of the safety envelope according to the lateral offset trajectory, and determining the roll angle command correction value based on the lateral offset trajectory under the constraints of the safety envelope, and correcting the lateral motion state of the vehicle according to the roll angle command correction value; After the correction, a safety boundary check is performed. If the test result does not fall within the safety envelope, the correction amount is recalculated and the correction operation is repeated. If it meets the safety envelope requirements, continuous monitoring is performed.

2. The flight trajectory control method based on operation state perception and instruction learning according to claim 1, characterized in that: The method of constructing a joint input tensor based on the operation sequence and the reference roll angle instruction sequence and obtaining a hidden state sequence after being processed by the encoder is as follows: The operation sequence and the reference roll angle command sequence are processed using a learnable embedding function, mapped into the same semantic space, and the two obtained embedding vectors are concatenated in dimension to form a joint input tensor; The joint input tensor is used as an input sequence, and after being processed by an encoder for extracting temporal features, a hidden state sequence is obtained, where the hidden state sequence contains context information and time dependency of the input sequence.

3. The flight trajectory control method based on operation state perception and instruction learning according to claim 1, characterized in that: The method for determining the predicted roll angle instruction is as follows: Based on the hidden state sequence, obtaining an operation probability distribution through a prediction model; The operation type with the highest probability under the operation probability distribution is used as the predicted editing operation type of the current time step; Based on the predicted editing operation type and in combination with a preset operation set, a roll angle instruction is determined.

4. The flight trajectory control method based on operation state perception and instruction learning according to claim 3 is characterized in that: Based on the predicted editing operation type and in combination with a preset operation set, a roll angle instruction is determined in the following manner: Construct an operation set based on historical data: ； in, is the correction amount obtained by clustering historical data. k Roll angle history correction, 、 、 、 Represents different operation types. Indicates that the base roll angle command remains unchanged. Indicates adding a k Class correction amount, Indicates that the kth type of correction is reduced based on the baseline instruction. Indicates direct replacement of the base instruction; k =1....K, K is the number of clusters; Determine the roll angle instruction based on the correspondence between the measurement editing operation type and the operation set : ； ； in, is the reference roll angle command, For offset compensation of REPLACE operation type, is hidden state, To map high-dimensional hidden features into linear weights of roll angle commands, To predict the type of editing operation, is the probability distribution, To find the editing operation type with the highest probability under the operation probability distribution.

5. The flight trajectory control method based on operation state perception and instruction learning according to claim 4, characterized in that: Domain knowledge constraints are imposed on the predicted editing operation type: ； in, t is the time step, T is the total number of time steps, and the operation distribution is 0 for the last 5 time steps of the sequence; is the probability distribution, Indicates a direct replacement of the base instruction.

6. The flight trajectory control method based on operation state perception and instruction learning according to claim 1, characterized in that: The method for obtaining the lateral offset trajectory is as follows: The kinematic model uses the roll angle command as input and calculates the basic lateral deviation trajectory of the carrier in the undisturbed state through the preset kinematic equation. ; ； The wind disturbance model is based on the temporal characteristics of the wind disturbance sequence and outputs the disturbance compensation for the lateral displacement of the carrier by fitting the wind field action law. ; ； The residual correction network obtains the residual correction by learning the model error law in the historical trajectory data ; ； ； Based on the basic trajectory, disturbance compensation, and residual correction, the corrected lateral deviation trajectory is obtained. : ； in, is the airspeed constant, is the time step, is the wind direction angle, is the cosine of the wind direction angle, is the predicted roll angle command of the current time step after sinusoidalization, is the predicted lateral offset of the previous time step, is the predicted roll angle command for the current time step, is the instantaneous wind speed and volume of the wind disturbance at the current time step, is the encoder output hidden layer dimension, is the input feature vector of the residual correction network; MLP is a multi-layer perceptron, are the learnable parameters in MLP, is hidden state, is the set of real numbers.

7. The flight trajectory control method based on operation state perception and instruction learning according to claim 6, characterized in that: The loss function of the lateral deviation trajectory for: ； ； ； ； in, is the editing operation loss, is the trajectory loss, is the wind disturbance consistency regularization term; is the wind disturbance impact weight, is the instantaneous wind speed and volume of the wind disturbance at the current time step, is the weight coefficient, i =1, 2, 3; For real editing operations, t is the time step, is the total number of time steps, is the probability distribution, is the residual correction, For time t The lateral offset trajectory, For time t The true offset trajectory, For time t Ideal lateral deviation trajectory under no wind conditions.

8. The flight trajectory control method based on operation state perception and instruction learning according to claim 6, characterized in that: The range of the safety envelope is adjusted according to the lateral deviation trajectory. The method is as follows: Determine the deviation trend and acceleration based on the initial heading deviation trajectory; Based on the excursion trend, acceleration, wind speed and corresponding flight phase, a safety envelope is constructed in combination with the preset excursion safety threshold. A predicted heading deviation trajectory is obtained according to the data received in real time, and the safety envelope is readjusted based on the predicted heading deviation trajectory.

9. The flight trajectory control method based on operation state perception and instruction learning according to claim 8, characterized in that: The flight data also includes the current lateral deviation trajectory of the vehicle, flight speed, deviation safety threshold, and envelope boundary slope. Under the safety envelope constraint, the roll angle command correction amount is determined based on the lateral deviation trajectory, and the method is as follows: Based on the lateral offset trajectory, the roll angle command correction sequence is obtained by reverse calculation; Based on the genetic algorithm, the roll angle command correction sequence is encoded as an individual of the genetic algorithm. The population iteration is combined with the fitness function to screen out the roll angle command correction sequence corresponding to the best individual as the candidate optimal solution. Based on the Lagrange multiplier method, with the minimization of the pilot's operation amount as the objective function and the safety envelope as the constraint condition, the optimal correction amount among the candidate optimal solutions is solved as the roll angle command correction amount.

10. A flight trajectory control system based on operation state perception and command learning, characterized in that the system includes: a data acquisition module configured to collect flight data in real time, the flight data including a current lateral deviation trajectory, flight speed, deviation safety threshold, envelope boundary slope, wind disturbance sequence, operation sequence, and reference roll angle command sequence; a preprocessing module configured to construct a joint input tensor based on the operation sequence and the reference roll angle command sequence, and obtain a hidden state sequence after being processed by the encoder; a prediction module configured to input the hidden state sequence into a preset prediction model to obtain an operation probability distribution, and then determine a roll angle command based on the operation probability distribution; an offset analysis module configured to construct a kinematic model based on the roll angle command to obtain a basic trajectory, construct a wind disturbance model based on the wind disturbance sequence to obtain a disturbance compensation amount, and correct a composite result of the basic trajectory and the disturbance compensation amount through a residual correction network to obtain a lateral offset trajectory; A safety envelope adjustment module is configured to analyze a deviation trend according to a lateral deviation trajectory and adjust a range of a safety envelope; a correction module configured to determine a roll angle command correction value based on the lateral offset trajectory under the constraints of the safety envelope, and correct the lateral motion state of the carrier according to the roll angle command correction value; The feedback module is configured to perform safety boundary detection after correction. If the detection result does not fall within the safety envelope, it returns to recalculate the correction amount and repeat the correction operation. If it meets the safety envelope requirements, it continues to monitor.

Images