Flight control optimization method, system, electronic equipment and storage medium

Abstract

The present invention relates to the field of flight control technology, and discloses a flight control optimization method, system, electronic device, and storage medium, including acquiring flight status data, performing data standardization and integrity processing; constructing an axial attention dynamic model to establish a state transition relationship; configuring a deep reinforcement learning framework and defining a multi-objective optimization mechanism; responding to the processed flight status data, based on the axial attention dynamic model and the deep reinforcement learning framework, iteratively updating network parameters through an interactive training process to achieve gradient optimization of the flight control strategy; responding to the trained flight status data, outputting a dynamic association model of environmental factors and control parameters, and generating a flight adjustment instruction set for deployment. The present invention improves modeling efficiency and reduces computational complexity through the fusion architecture of axial attention and deep reinforcement learning, while having strong generalization capabilities, ultimately improving control accuracy and enhancing flight safety.

Description

Technical Field

[0001] The present invention belongs to the field of flight control, and in particular relates to a flight control optimization method, system, electronic equipment and storage medium. Background Art

[0002] Currently, flight control optimization technology has the following main limitations:

[0003] Insufficient environmental adaptability:

[0004] Traditional PID control and linear regression models rely on fixed parameter rule bases and are unable to dynamically respond to complex environmental changes such as sudden changes in wind speed and sudden drops in visibility, resulting in increased control deviation.

[0005] Timing modeling is inefficient:

[0006] While LSTM-based prediction models can process time series data, they struggle to effectively capture the high-dimensional coupling relationships between operational instructions, posture parameters, and environmental factors in long sequences. When the time step size exceeds 50, model training time increases exponentially, and prediction accuracy decreases by over 15%.

[0007] The strategy optimization dimension is single:

[0008] Existing reinforcement learning methods mostly focus on a single control objective (such as posture stability) and have not established a dynamic association mechanism between environmental risks and operating parameters.

[0009] Missing safety constraints:

[0010] Mainstream deep reinforcement learning frameworks (such as DQN and PPO) lack built-in safety mechanisms and are prone to generating high-risk instructions under extreme environmental conditions.

[0011] Therefore, the present application provides a flight control optimization method to solve the above technical problems. Summary of the Invention

[0012] The purpose of the present invention is to provide a flight control optimization method, system, electronic device and storage medium to solve the technical problems in the prior art such as low modeling efficiency, complex calculation and poor generalization ability, which lead to reduced control accuracy and safety.

[0013] In order to solve the above technical problems, the present invention provides a flight control optimization method, comprising:

[0014] performing data normalization and integrity processing in response to the acquired flight status data, wherein the flight status data includes original time series data sets of pilot operation instructions, aircraft attitude parameters, and environmental monitoring indicators;

[0015] Constructing an axial attention dynamic model to establish state transition relationships, including constructing an axial transformer network based on a dual-path attention mechanism to learn the spatiotemporal dependencies of flight state transitions;

[0016] Configure the deep reinforcement learning framework and define the multi-objective optimization mechanism, including the state space definition, action space mapping, and reward mechanism of the multi-objective optimization learning framework;

[0017] In response to the processed flight state data, based on the axial attention dynamic model and the deep reinforcement learning framework, iteratively updating network parameters through an interactive training process to achieve gradient optimization of the flight control strategy;

[0018] In response to the flight status data after training, a dynamic correlation model between environmental factors and control parameters is output, and a flight adjustment instruction set for deployment is generated.

[0019] In some specific embodiments, data normalization and integrity processing are performed in response to acquired flight status data, wherein the flight status data includes original time series data sets of pilot operation instructions, aircraft attitude parameters, and environmental monitoring indicators, further comprising:

[0020] Obtaining the pilot's operating instructions, including obtaining physical quantities of the operating instructions, obtaining the aircraft's attitude parameters, including obtaining attitude angle parameters, and obtaining the environmental monitoring indicators, including obtaining atmospheric environment monitoring data;

[0021] performing differential normalization processing on the flight status data, including performing a first type of range compression transformation on the pilot operation command data and a second type of symmetric range transformation on the aircraft attitude parameter data, wherein the first type of range compression transformation is a unit dimension interval and the second type of symmetric range transformation is a symmetric interval;

[0022] In response to the flight status data after differential normalization, a time series interpolation algorithm is used to compensate for missing data segments, and outlier data points are identified and removed based on a probability distribution model;

[0023] Construct a structured dataset, which involves splitting continuous time series data into training, validation, and test subsets with fixed time windows.

[0024] In some specific embodiments, constructing an axial attention dynamic model to establish state transition relationships includes constructing an axial transformer network based on a dual-path attention mechanism for learning spatiotemporal dependencies of flight state transitions, further comprising:

[0025] The axial attention dynamic model includes a time dimension attention layer and a feature dimension attention layer. The time dimension attention layer captures the long-term dependency features in the operation sequence and generates a time dimension feature representation. The feature dimension attention layer learns the coupling relationship between different flight status data and constructs a multi-dimensional feature interaction matrix.

[0026] Configuring independently adjusted time dimension and feature dimension multi-head attention components to enhance the representation capability of the axial attention dynamic model;

[0027] Construct a hierarchical network architecture with residual connections to prevent degradation during training.

[0028] In some specific embodiments, configuring a deep reinforcement learning framework and defining a multi-objective optimization mechanism, including a multi-objective optimization learning framework for state space definition, action space mapping, and reward mechanism, further includes:

[0029] The state space definition includes constructing a multi-dimensional state vector of operational control variables, flight attitude variables, and environmental disturbance variables;

[0030] The motion space mapping includes defining a continuous motion output space reflecting thrust adjustment instructions and attitude correction instructions;

[0031] The reward mechanism includes constructing a penalty function based on the target state deviation measurement and introducing environmental safety boundary constraints;

[0032] Deploy a random strategy generation network for action decision-making and simultaneously build a value evaluation network.

[0033] In some specific embodiments, in response to the processed flight state data, based on the axial attention dynamic model and the deep reinforcement learning framework, iteratively updating network parameters through an interactive training process to achieve gradient optimization of the flight control strategy further includes:

[0034] Initialize network parameters according to the differentiation strategy, including using an orthogonal matrix to initialize strategy network parameters and initializing value assessment network parameters based on the variance stability principle;

[0035] Construct a loss function based on the value assessment network, including constructing a mean square error loss function of the temporal difference error, to optimize the prediction accuracy of the value assessment network;

[0036] Set up an adaptive learning rate mechanism, including implementing a phased decay learning rate adjustment mechanism to balance training speed and convergence stability;

[0037] A priority sampling strategy is constructed based on the sample importance weights to accelerate the learning process of key samples.

[0038] In some specific embodiments, in response to the flight status data after training, outputting a dynamic association model between environmental factors and control parameters and generating a flight adjustment instruction set for deployment further includes:

[0039] Calculate the real-time influence coefficient matrix of environmental interference parameters and flight control parameters;

[0040] Generate a three-dimensional dynamic surface map reflecting the environment-operation relationship;

[0041] Dynamically output thrust compensation coefficient and attitude conservative adjustment amount according to environmental parameter risk level;

[0042] The strategy robustness is tested in a flight simulation environment, and trajectory tracking accuracy and energy consumption indicators are quantitatively evaluated.

[0043] In some specific embodiments, the deep reinforcement learning framework further includes:

[0044] A stabilization training mechanism for target network separation, including the use of dual-target network architecture separation strategy improvement and value estimation process;

[0045] Integrate environmental over-limit penalty terms and safe operation reward terms into the basic reward function;

[0046] Inject parameterized noise signals into the action selection layer to expand the strategy search space;

[0047] Configure a discount factor close to unity to balance immediate operational benefits with long-term flight objectives.

[0048] Based on the same concept, the present invention also provides a flight control optimization system, comprising:

[0049] a data acquisition module configured to perform data normalization and integrity processing in response to acquired flight status data, wherein the flight status data includes original time series data sets of pilot operation instructions, aircraft attitude parameters, and environmental monitoring indicators;

[0050] a dynamic model building module configured to build an axial attention dynamic model to establish a state transition relationship, including building an axial transformer network based on a dual-path attention mechanism for learning the spatiotemporal dependency of flight state transitions;

[0051] A learning framework establishment module is configured to configure a deep reinforcement learning framework and define a multi-objective optimization mechanism, including a multi-objective optimization learning framework for state space definition, action space mapping, and reward mechanism;

[0052] a data training module configured to iteratively update network parameters through an interactive training process in response to the processed flight state data based on the axial attention dynamic model and the deep reinforcement learning framework to achieve gradient optimization of the flight control strategy;

[0053] The flight control optimization module is configured to output a dynamic association model between environmental factors and control parameters in response to the flight status data after training, and generate a flight adjustment instruction set for deployment.

[0054] Based on the same concept, the present invention also provides an electronic device, including: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of a flight control optimization method.

[0055] Based on the same concept, the present invention also provides a computer-readable storage medium storing a computer program executable by an electronic device. When the computer program runs on the electronic device, the electronic device executes the steps of a flight control optimization method.

[0056] Compared with the prior art, the beneficial effects are:

[0057] The present invention discloses a flight control optimization method, system, electronic device and storage medium. Through the fusion architecture of axial attention and deep reinforcement learning, modeling efficiency is improved, computational complexity is reduced, and generalization capability is strong, ultimately improving control accuracy and enhancing flight safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] 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:

[0059] Figure 1 This is a flow chart of a flight control optimization method according to some specific embodiments of the present invention;

[0060] Figure 2 This is one of the flowcharts of a flight control optimization method of the present invention in some applications;

[0061] Figure 3 This is the second flow chart of a flight control optimization method of the present invention in some applications;

[0062] Figure 4 This is the third flow chart of a flight control optimization method of the present invention in some applications;

[0063] Figure 5 This is a fourth flow chart of a flight control optimization method of the present invention in some applications;

[0064] Figure 6 This is the fifth flow chart of a flight control optimization method of the present invention in some applications;

[0065] Figure 7 This is the sixth flow chart of a flight control optimization method of the present invention in some applications;

[0066] Figure 8 Schematic diagram of the structure of a flight control optimization system of the present invention in some specific embodiments;

[0067] Figure 9 It is a schematic structural diagram of an electronic device according to some specific embodiments of the present invention.

[0068] In the figure, 710 is a processor; 720 is a memory; 730 is an input device; and 740 is an output device. DETAILED DESCRIPTION

[0069] To make the objectives, technical solutions, and advantages of this application more clear, this application will be further described in detail below with reference to the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this application.

[0070] The terms used in the examples of this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The singular forms "a," "the," and "the" used in the examples of this application and the appended claims are also intended to include plural forms, and unless the context clearly indicates otherwise, "a plurality" generally includes at least two.

[0071] It should be understood that the term "and / or" as used herein is merely a description of the relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the associated objects are in an "or" relationship.

[0072] It should be understood that although the terms first, second, third, etc. may be used to describe in the embodiments of the present application, these descriptions should not be limited to these terms. These terms are only used to distinguish the descriptions. For example, without departing from the scope of the embodiments of the present application, the first may also be referred to as the second, and similarly, the second may also be referred to as the first.

[0073] As used herein, the words "if" and "if" may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to the determination" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)," depending on the context.

[0074] It should also be noted that the terms "include," "comprises," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a product or device comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such product or device. In the absence of further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the product or device comprising the element.

[0075] It should be noted in particular that any symbols and / or numbers in the specification that are not marked in the accompanying drawings are not drawing marks.

[0076] Reference Figure 1 , a flight control optimization method, comprising:

[0077] S101, performing data standardization and integrity processing in response to acquired flight status data, wherein the flight status data includes original time series data sets of pilot operation instructions, aircraft attitude parameters, and environmental monitoring indicators;

[0078] S102, constructing an axial attention dynamic model to establish state transition relationships, including constructing an axial transformer network based on a dual-path attention mechanism to learn the spatiotemporal dependencies of flight state transitions;

[0079] S103, configuring a deep reinforcement learning framework and defining a multi-objective optimization mechanism, including a multi-objective optimization learning framework for state space definition, action space mapping, and reward mechanism;

[0080] S104, in response to the processed flight status data, iteratively updating network parameters through an interactive training process based on the axial attention dynamic model and the deep reinforcement learning framework to achieve gradient optimization of the flight control strategy;

[0081] S105 , in response to the flight status data after training, output a dynamic association model between environmental factors and control parameters, and generate a flight adjustment instruction set for deployment.

[0082] Specifically, in an embodiment of the present invention, the original flight status dataset obtained from the onboard sensors and flight data recorder is first responded to. The dataset continuously collects the operating energy corresponding to the pilot's thrust control stick displacement instructions, the aircraft's three-axis attitude angle parameters, and the wind speed and fog concentration indicators fed back in real time by the atmospheric environment monitoring unit, forming a time series data stream with millisecond-level accuracy; to eliminate the impact of dimensional differences on subsequent modeling, a first type of range compression transformation is performed on the operating energy data, such as mapping the throttle stick displacement percentage to a unit interval, and a second type of symmetric range transformation is performed on the roll angle attitude parameter, such as normalizing the angle value to a standard interval symmetric about zero; to address the signal loss problem that may occur in data transmission, a sequence compensation algorithm based on linear interpolation of previous and next time points is used to repair the missing segments, and abnormal pulse data points are identified and eliminated based on the triple standard deviation principle. Finally, a structured dataset consisting of a training subset, a validation subset, and a test subset is constructed, and each subset maintains a continuous time window characteristic to avoid time series breaks.

[0083] Subsequently, an axial attention dynamic model is constructed to establish a state transfer relationship. In an embodiment of the present invention, the input sequence is processed in the time dimension attention layer, and the long-term dependency characteristics of the operation energy evolving over time are extracted through attention weight calculation, such as capturing the inertial delay effect of the thrust adjustment instruction when encountering continuous wind shear; the coupling relationship between different parameters is analyzed in the feature dimension attention layer, such as establishing a correlation matrix between the roll angle change and the side wind speed disturbance; the model adopts an independently configurable multi-head attention mechanism, and the number of attention heads in the time axis and feature axis are both set to integers greater than eight to enhance the feature extraction capability, and at the same time, a residual connection module is embedded to prevent gradient disappearance, such as transferring the original posture parameter features in the deep layer of the network to maintain information integrity.

[0084] Next, a multi-objective optimization mechanism is configured within the deep reinforcement learning framework: the state space is defined as a four-dimensional vector, containing normalized thrust operation energy, roll angle attitude parameters, wind speed, and fog concentration monitoring values; the action space is mapped into a two-dimensional continuous vector, corresponding to the thrust adjustment percentage increment and the roll angle correction angle increment, respectively; a composite reward function is constructed. The basic part calculates the absolute deviation between the current thrust and the target thrust, weighted by the roll angle deviation. The weight coefficient is dynamically adjusted according to the flight phase, for example, giving the attitude angle a higher weight during landing. At the same time, an environmental safety constraint is introduced, imposing a negative penalty when the wind speed exceeds the set risk threshold. For example, an additional penalty is added for each unit increase in wind speed above eight meters per second.

[0085] Based on the aforementioned architecture, interactive training is initiated in response to preprocessed flight status data: orthogonal initialization of policy network parameters is adopted, and variance-stabilized initialization of value assessment network parameters is adopted; each round of training samples are transferred from the state of the priority experience replay pool, and the sample weights are dynamically adjusted by the historical time series difference error, for example, a three-fold sampling probability is given to attitude out-of-control samples; the network parameters are updated by the gradient descent method, and the initial learning rate is set to one thousandth and exponentially decayed in stages; parameterized Gaussian noise is injected into the action selection layer to enhance exploration, such as superimposing a random perturbation with a standard deviation of five percent on the thrust adjustment instruction; after ten thousand iterations, the policy network can generate stable control instructions, such as automatically increasing the thrust baseline value by five percent under level six wind speed conditions.

[0086] After training, a dynamic correlation model between environmental factors and control parameters is output: the influence coefficients of wind speed and thrust operation are calculated in real time, for example, a linear coefficient table is generated for the percentage of thrust compensation required for every meter per second increase in wind speed; the nonlinear relationship between fog concentration and roll angle correction is visualized through a three-dimensional surface graph; an adaptive instruction set is generated when the environmental risk level increases, for example, a conservative strategy is activated when the fog concentration exceeds 50%, and the maximum roll angle correction is compressed to 80% of the standard value; finally, it is deployed in a flight simulator for closed-loop verification. In the scenario of reproducing historical wind shear events, the system automatically increases the thrust by 12% and reduces the roll angle by 3 degrees, reducing the trajectory deviation distance by more than 60% and reducing fuel consumption by 15%.

[0087] like Figure 2 As shown, in some of the applications, data standardization and integrity processing are performed in response to the acquired flight status data, wherein the flight status data includes original time series data sets of pilot operation instructions, aircraft attitude parameters and environmental monitoring indicators, including obtaining the pilot operation instructions, including obtaining the physical quantity of the operation instructions, obtaining the aircraft attitude parameters, including obtaining the attitude angle parameters, obtaining the environmental monitoring indicators, including obtaining atmospheric environment monitoring data; performing differential normalization processing on the flight status data, including performing a first type of range compression transformation on the pilot operation instruction data, and performing a second type of symmetric range transformation on the aircraft attitude parameter data, wherein the first type of range compression transformation is a unit dimension interval, and the second type of symmetric range transformation is a symmetric interval; in response to the flight status data after differential normalization processing, a time series interpolation algorithm is used to compensate for missing data segments, and outlier data points are identified and eliminated based on a probability distribution model; and constructing a structured data set, including dividing the continuous time series data into a training subset, a validation subset and a test subset with fixed time windows.

[0088] It can be understood that in this application, in response to the acquired original flight status time series data set, multi-source heterogeneous data are first collected synchronously: the physical value corresponding to the pilot's thrust operation instruction is captured through the joystick displacement sensor, such as recording the continuous displacement signal of the throttle lever in the range of zero to maximum travel; the three-axis attitude angle parameters of the aircraft are obtained through the attitude reference system, such as the real-time change trajectory of the roll angle and the pitch angle; the atmospheric data computer is connected to collect environmental monitoring indicators, such as the fog concentration value converted from the forward wind speed and atmospheric transmittance; in order to eliminate the interference of different physical dimensions on the model training, the first type of range compression transformation is performed on the operation instruction data, such as linearly mapping the original voltage signal of the throttle lever displacement to the unit dimension interval, so that the minimum displacement corresponds to the lower limit of the transformation and the maximum displacement corresponds to the upper limit of the transformation; at the same time, the second type of symmetry range is performed on the attitude angle parameters. For example, the original roll angle is mapped to a standard symmetrical interval with zero angle as the symmetry center, so that the positive and negative maximum deflection angles correspond to the endpoints of the interval respectively. After completing the differential normalization processing, in order to solve the problem of instantaneous loss that may occur in data collection, a time series interpolation algorithm based on the weighted average of adjacent time points is used to compensate for the missing segments. For example, when the signal of a millisecond-level sampling point is lost, the arithmetic mean of the two valid sampling points before and after is automatically taken to fill the gap. Outliers are identified based on the probability distribution model of historical data. For example, the standard deviation of the attitude angle in 100 consecutive sampling windows is calculated, and data points exceeding the range of three times the standard deviation are eliminated. Finally, a structured data set is constructed, and the continuous time series stream is divided into fixed time windows, such as slicing every five seconds, and distributed to the training subset, validation subset and test subset according to the preset ratio to ensure that each subset contains the complete flight phase characteristics.

[0089] like Figure 3 As shown, in some of the applications, an axial attention dynamic model is constructed to establish a state transfer relationship, including constructing an axial transformer network based on a dual-path attention mechanism for learning the spatiotemporal dependency of flight state transfer, including the axial attention dynamic model, including a time dimension attention layer and a feature dimension attention layer, capturing long-term dependency features in the operation sequence through the time dimension attention layer, generating a time dimension feature representation, learning the coupling relationship between different flight state data in the feature dimension attention layer, and constructing a multi-dimensional feature interaction matrix; configuring independently adjusted time dimension and feature dimension multi-head attention components to enhance the representation capability of the axial attention dynamic model; constructing a hierarchical network architecture containing residual connections to prevent the training process from degenerating.

[0090] It can be understood that in this application, in the stage of constructing the axial attention dynamic model, the axial transformer network architecture of the dual-path attention mechanism is first deployed: the time dimension attention layer receives the preprocessed time series data stream, and captures the long-term dependency features in the operation sequence through the sliding window mechanism, such as analyzing the inertial delay effect of continuous thrust adjustment instructions when encountering continuous wind shear, and generating a feature representation with time correlation; the feature dimension attention layer synchronously processes the time feature output, learns the coupling relationship between different flight state parameters, such as establishing a nonlinear correlation matrix between roll angle changes and lateral wind speed disturbances, and constructing a multi-dimensional feature interaction that reflects the attitude-environment interaction mechanism. Model; To enhance the model's representation capabilities, independently adjustable multi-head attention components are configured. The time dimension head group focuses on the evolution pattern recognition of operation instructions. For example, a time dimension attention head of no less than a preset base number is set to extract the frequency domain features of thrust regulation. The feature dimension head group specializes in parameter correlation mining. For example, independent head groups are configured to learn the weights of the influence of wind speed on roll angle and fog concentration on pitch angle respectively. A hierarchical architecture with residual connections is constructed in the deep layer of the network. For example, a cross-layer direct connection is introduced after the fifth attention layer to bypass the original posture parameter features to the downstream module to prevent the training degradation problem caused by gradient disappearance and ensure the stability of state transfer relationship modeling in extreme environments.

[0091] like Figure 4 As shown, in some of the applications, a deep reinforcement learning framework is configured and a multi-objective optimization mechanism is defined, including a multi-objective optimization learning framework of state space definition, action space mapping and reward mechanism, including that the state space definition includes constructing a multi-dimensional state vector of operation control variables, flight attitude variables and environmental interference variables; the action space mapping includes defining a continuous action output space reflecting thrust adjustment instructions and attitude correction instructions; the reward mechanism includes constructing a penalty function based on the target state deviation measurement and introducing environmental safety boundary constraints; deploying a random strategy generation network for action decision-making, and synchronously constructing a value evaluation network.

[0092] It is understandable that in this application, during the deep reinforcement learning framework configuration phase, a multidimensional state space is first constructed: pre-processed operational control variables (such as normalized thrust control), flight attitude variables (such as symmetric roll angle parameters), and environmental interference variables (such as wind speed and fog concentration monitoring values) are integrated to form a four-dimensional vector representing the complete flight state. For example, during the approach phase, a thrust value of 0.85, a roll angle of -0.3, a wind speed of 0.6, and a fog concentration of 0.4 are encoded as the state vector [0.85, -0.3, 0.6, 0.4]. A continuous action output space mapping mechanism is defined, with thrust adjustment instructions using unit adjustment amounts to represent the direction and magnitude of increase or decrease (such as +0.05 to increase the standard thrust value by 5%), and attitude correction instructions using standard angle increments (such as -0.1 to reduce the roll angle by 1.5 degrees). A multi-objective reward function is designed. The basic penalty term calculates the absolute deviation between the actual thrust value and the target value, weighted by the attitude angle deviation (for example, the attitude angle weight is increased to 0.7 times during landing). Environmental safety boundary constraints are simultaneously introduced. When the wind speed exceeds the set risk threshold, a dynamic penalty term is added (for example, a double penalty is imposed for each unit increase in wind speed after the wind speed exceeds the standard threshold; for example, during cruise phase: thrust weight W1=0.6, attitude weight W2=0.3; during landing phase: W1=0.3, W2=0.7). A random strategy is deployed to generate a network output action probability distribution. For example, the probability of generating an adjustment command of [+0.08,-0.05] under level 6 wind speed conditions is 70%. A value assessment network is also constructed. For example, the expected long-term cumulative reward for executing the action [+0.05,0] under the state [thrust 0.85, wind speed 0.6] is calculated.

[0093] like Figure 5 As shown, in some of the applications, in response to the processed flight status data, based on the axial attention dynamic model and the deep reinforcement learning framework, the network parameters are iteratively updated through an interactive training process to achieve gradient optimization of the flight control strategy, including initializing the network parameters according to the differentiation strategy, including using an orthogonal matrix to initialize the strategy network parameters, and initializing the value evaluation network parameters based on the variance stability principle; constructing a loss function based on the value evaluation network, including constructing a mean square error loss function of the time difference error to optimize the prediction accuracy of the value evaluation network; setting an adaptive adjustment learning rate mechanism, including implementing a staged attenuation learning rate adjustment mechanism to balance the training speed and convergence stability; constructing a priority sampling strategy based on the sample importance weight to accelerate the learning process of key samples.

[0094] As can be understood, in this application, interactive training of the axial attention model and deep reinforcement learning is initiated in response to preprocessed flight state time series data. First, a parameter initialization strategy is implemented based on network functional differences. The policy network uses an orthogonal matrix initialization method, for example, constructing an orthogonal basis vector set in the policy output layer to ensure a uniform distribution across the action space. The value network is initialized based on the variance stability principle, for example, by adjusting the weights of the fully connected layer to keep the standard deviation of its output within a preset range. A value-based loss function is constructed, centered on the mean squared error of the temporal difference error (MSE), for example, comparing the squared deviation between the value network's prediction of the current state action and the target network's estimate of the maximum value of the next state. An adaptive learning rate mechanism with staged decay is implemented. Initially, a higher learning rate is used to accelerate convergence, for example, an initial value of one thousandth. When the reward gain for three consecutive training cycles falls below a preset threshold, a stepwise decay is triggered, for example, by 50% of the original value each time, balancing convergence speed and stability. A priority sampling strategy is constructed based on sample importance weights. The absolute value of the time series difference error (TDE) of each state transition sample is calculated as a priority benchmark. For example, samples of attitude loss (roll angle deviation exceeding the safety margin) are given a sampling weight tripled. These key samples are prioritized in the experience replay pool to accelerate policy optimization. After multiple rounds of iteration, training is terminated when the average reward value on the validation set has not significantly improved after 100 consecutive updates. At this point, the policy network can generate stable instructions. For example, in scenarios with a sudden increase in fog concentration, the thrust compensation mechanism is automatically triggered, reducing flight trajectory deviation by over 60%.

[0095] like Figure 6 As shown, in some of the applications, in response to the flight status data after training, a dynamic correlation model of environmental factors and control parameters is output, and a flight adjustment instruction set for deployment is generated, including calculating a real-time influence coefficient matrix of environmental interference parameters and flight control parameters; generating a three-dimensional dynamic surface map reflecting the environment-operation correlation relationship; dynamically outputting the thrust compensation coefficient and attitude conservative adjustment amount according to the risk level of the environmental parameters; testing the robustness of the strategy in a flight simulation environment, and quantitatively evaluating the trajectory tracking accuracy and energy consumption indicators.

[0096] It can be understood that in this application, in response to the flight status data after training, the dynamic influence coefficient matrix between the environmental interference parameters and the flight control parameters is calculated in real time, for example, a linear response coefficient table of the unit change of wind speed to the thrust operation amount is constructed, and the reference value of the roll angle compensation corresponding to each 10% increase in fog concentration is recorded; based on this matrix, a three-dimensional dynamic surface map is generated, the horizontal axis maps the environmental parameters (such as wind speed 0 to the maximum monitoring value), the vertical axis represents the control parameters (such as thrust compensation coefficient 0 to 0.2), the surface height indicates the adjustment efficiency (such as the improvement degree of trajectory stability), and the environment-operation relationship is displayed through an interactive visual interface, for example, the optimal thrust compensation range is displayed at the intersection of a wind speed of eight meters per second and a fog concentration of thirty percent. According to the dynamic environmental risk classification mechanism, when the wind speed exceeds the preset safety boundary, the forward thrust compensation coefficient is automatically output (for example, an additional 2% of the baseline thrust is added for every meter per second the wind speed exceeds the standard). When the fog concentration enters the low visibility range, the conservative attitude adjustment is activated (for example, when the concentration exceeds 50%, the maximum roll angle correction is compressed to 80% of the standard value). Finally, historical wind shear event data streams are injected into a high-fidelity flight simulation environment to test the robustness of the strategy: trajectory tracking accuracy indicators are quantified (for example, the root mean square deviation between the actual track and the planned route is reduced by more than 60%), and energy consumption indicators are simultaneously monitored (for example, the fuel flow meter shows a 15% reduction in consumption rate in turbulent scenarios) to verify the effectiveness of the instruction set in extreme environments.

[0097] like Figure 7 As shown in the figure, in some of these applications, the deep reinforcement learning framework also includes a stabilized training mechanism with separated target networks, including the use of a dual-target network architecture to separate the policy improvement and value estimation processes; integrating environmental limit penalty terms and safe operation reward terms in the basic reward function; injecting parameterized noise signals into the action selection layer to expand the policy search space; and configuring a discount factor close to unity to balance the immediate operation benefits and long-term flight goals.

[0098] As can be understood, in this application, the deep reinforcement learning framework first deploys a stabilization training mechanism with a separate objective network. A dual-objective network architecture is employed to separate the policy improvement and value estimation processes. The policy optimization network generates action commands in real time, while the value benchmark network periodically synchronizes parameters to provide stable target values. For example, in wind shear scenario training, the value benchmark network synchronizes parameters every 100 iterations, reducing training fluctuations by 70%. A composite optimization term is integrated into the basic reward function. In addition to the state tracking deviation penalty, it also adds an environmental overrun penalty (such as an exponentially increasing dynamic penalty when wind speed exceeds the safety margin) and a safe operation reward (such as an additional positive reward for maintaining attitude stability for ten seconds). For example, in a fog concentration gradient increasing scenario, this composite reward mechanism increased the safe operation rate by 40%. A parameterized noise signal is injected into the action selection layer to expand the policy search space. Controllable random perturbations are superimposed on the output continuous actions. For example, the introduction of Gaussian noise with a standard deviation of 5% of the command amplitude into the thrust adjustment command triples the policy's coverage in turbulence exploration. Configuring a discount factor close to unity to balance the decision-making horizon, for example, setting an attenuation coefficient of 0.98, allows the system to prioritize immediate track accuracy during the approach phase while taking into account long-term fuel economy goals. Simulation verification shows that this configuration reduces fuel consumption by 12 percentage points during a 5,000-meter approach segment.

[0099] In another embodiment of the present invention, an Axial Transformer combined with Deep Reinforcement Learning (DRL) is employed. By decomposing the attention mechanism, the Axial Transformer can efficiently process long sequences of data while maintaining the ability to dynamically model time series. DRL is then used to optimize the pilot's operational strategy to maximize flight performance and safety.

[0100] Model representation:

[0101] State space representation: define the state vector for:

[0102] in:

[0103] : The pilot inputs the operating energy (such as thrust percentage);

[0104] : attitude slope performance (such as roll angle);

[0105] : Environmental factors (such as wind speed, unit: m / s);

[0106] : Environmental factors (such as fog concentration, unit: visibility percentage).

[0107] Action space representation:

[0108] Defining the motion vector for:

[0109] in:

[0110] : The increment of the pilot's adjustment operation energy (unit: percentage);

[0111] : Adjust the increment of attitude slope performance (unit: degree).

[0112] Environmental feedback:

[0113] Define the reward function for:

[0114] in:

[0115] 、 is the weight coefficient, which is used to balance the importance of different state variables;

[0116] 、 is the target state value, such as target thrust percentage and target roll angle.

[0117] State transition function:

[0118] It can be learned through the Axial Transformer network, the specific form is:

[0119] ;

[0120] Detailed mathematical expression of Axial Transformer:

[0121] Axial Transforme decomposes the two-dimensional attention into two one-dimensional attentions, thereby reducing computational complexity. For time series data, one-dimensional attention operations can be performed on the time step and feature dimensions respectively.

[0122] Timeline Attention:

[0123] ;

[0124] ;

[0125] Feature axis attention:

[0126] ;

[0127] ;

[0128] Output:

[0129] ;

[0130] The optimization process of deep reinforcement learning:

[0131] In deep reinforcement learning, the goal of the agent (pilot) is to maximize the cumulative reward. and value network , the agent can learn the optimal action strategy.

[0132] Policy Network:

[0133] ;

[0134] in are the parameters of the policy network.

[0135] Value Network:

[0136] ;

[0137] in are the parameters of the value network, is the discount factor.

[0138] Loss function:

[0139] ;

[0140] Optimization process:

[0141] Use gradient descent to update the parameters of the policy network and value network:

[0142] ;

[0143] ;

[0144] in and is the learning rate.

[0145] Parameter Description:

[0146] Axial Transformer parameters:

[0147] Timeline attention heads : Select according to data complexity, such as 8 or 16.

[0148] Number of attention heads on feature axis : Select according to data complexity, such as 8 or 16.

[0149] Hidden layer dimensions : Can be set to 256 or 512.

[0150] Time step : Select according to the dynamic characteristics of the flight mission, such as 10 or 20.

[0151] Reward function weights:

[0152] 、 The equal weight coefficient is adjusted according to actual needs to balance the importance of different state variables.

[0153] Learning rate:

[0154] Policy network learning rate : Usually set to or .

[0155] Value network learning rate : Usually set to or .

[0156] This embodiment collects pilot operation data under different environmental conditions, including operation energy, attitude and slope energy, wind speed, fog concentration, etc. The data is normalized and the attention mechanism of time axis and feature axis is defined. Time axis attention is responsible for capturing the dynamic changes of time series, while feature axis attention is responsible for capturing the relationship between different features. Set the hidden layer dimension , number of attention heads and Parameters such as . Use the multi-head attention mechanism to improve the model's ability to model different subspaces. Define the policy network and value network The policy network is used to generate actions, and the value network is used to evaluate the value of actions. Set the reward function , according to the target state value and , calculate the reward value of the current state and action. Set the discount factor , used to balance the importance of immediate rewards and future rewards.

[0157] Randomly initialize the parameters of the Axial Transformer and DRL models. You can use Xavier initialization or He initialization to ensure that the initial values ​​of the parameters are reasonable. Use input data and Train the Axial Transformer. Learn the complex relationship between state and action using attention mechanisms across time and feature axes. Optimize the pilot's maneuver strategy using the policy network and value network of DRL. Update the policy network parameters using policy gradients, enabling the agent to select optimal actions. Update the value network parameters using gradient descent to minimize the error between predicted and target values. Use Experience Replay to store and replay historical data, improving model stability and convergence speed.

[0158] Evaluate model performance using the validation dataset to ensure it accurately predicts pilot maneuver and attitude and slope performance. Calculate metrics such as mean squared error (MSE) and mean absolute error (MAE) to assess model prediction accuracy. Adjust model parameters and optimize the model structure based on the validation results. For example, adjust parameters such as hidden layer dimensions, number of attention heads, and learning rate to improve model performance. Use early stopping techniques to prevent overfitting. When the error on the validation dataset stops decreasing, stop training and select the optimal model parameters. Analyze the relationship between the pilot maneuver and attitude and slope performance and environmental factors (such as wind and fog) using model output. For example, calculate the changing trends of maneuver and attitude and slope performance under different wind speeds and fog concentrations. Use visualization tools (such as Matplotlib or Seaborn) to plot these ratios to visually demonstrate the relationship between different variables. For example, plot a scatter plot of maneuver and wind speed to analyze the correlation between the two.

[0159] Based on the model results, the pilot's operational strategy is optimized. For example, when wind speeds are high, the pilot's operational capacity is appropriately increased to maintain flight stability; when fog concentration is high, the pilot's attitude and slope capacity are adjusted to avoid flight deviation. The optimized strategy is applied in a flight simulator or in real-world missions to verify its effectiveness and safety. The effectiveness of the optimized strategy is evaluated by comparing flight performance indicators before and after optimization.

[0160] For the method steps disclosed in the above embodiments, for the purpose of simple description, the method steps are expressed as a series of action combinations. However, those skilled in the art should be aware that the embodiments of the present invention are not limited by the order of the actions described, because according to the embodiments of the present invention, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

[0161] like Figure 8 As shown, the present invention also provides a flight control optimization system, comprising:

[0162] a data acquisition module 201 configured to perform data normalization and integrity processing in response to acquired flight status data, wherein the flight status data includes original time series data sets of pilot operation instructions, aircraft attitude parameters, and environmental monitoring indicators;

[0163] a dynamic model building module 202 configured to construct an axial attention dynamic model to establish a state transition relationship, including constructing an axial transformer network based on a dual-path attention mechanism for learning the spatiotemporal dependency of flight state transitions;

[0164] A learning framework establishment module 203 is configured to configure a deep reinforcement learning framework and define a multi-objective optimization mechanism, including a state space definition, an action space mapping, and a multi-objective optimization learning framework of a reward mechanism;

[0165] a data training module 204 configured to iteratively update network parameters through an interactive training process in response to the processed flight state data based on the axial attention dynamic model and the deep reinforcement learning framework to achieve gradient optimization of the flight control strategy;

[0166] The flight control optimization module 205 is configured to output a dynamic correlation model between environmental factors and control parameters in response to the flight status data after training, and generate a flight adjustment instruction set for deployment.

[0167] It is worth noting that although only some basic functional modules are disclosed in the embodiment of the present invention, it does not mean that the composition of the present system is limited to the above basic functional modules. On the contrary, what this embodiment wants to express is that on the basis of the above basic functional modules, those skilled in the art can arbitrarily add one or more functional modules in combination with the existing technology to form an infinite number of embodiments or technical solutions. In other words, this system is open rather than closed. Just because this embodiment only discloses individual basic functional modules, it cannot be considered that the scope of protection of the claims of the present invention is limited to the disclosed basic functional modules. At the same time, for the convenience of description, the above devices are described in terms of functions, which are divided into various units and modules. Of course, when implementing the present invention, the functions of each unit and module can be implemented in the same or one or more software and / or hardware.

[0168] like Figure 9 As shown, the present invention also provides an electronic device, comprising: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; a computer program is stored in the memory, and when the computer program is executed by the processor, the processor executes the steps of a flight control optimization method.

[0169] Figure 9Schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. Figure 9 The electronic device provided in the embodiment of the present invention includes: one or more processors 710 and a memory 720; the processor 710 in the electronic device may be one or more, Figure 9 Take a processor 710 as an example; the memory 720 is used to store one or more programs; the one or more programs are executed by the one or more processors 710, so that the one or more processors 710 implement a flight control optimization method as described in any one of the embodiments of the present invention.

[0170] The electronic device may further include an input device 730 and an output device 740 .

[0171] The processor 710, memory 720, input device 730 and output device 740 in the electronic device can be connected via a bus or other means. Figure 9 The bus connection is taken as an example.

[0172] Memory 720 in the electronic device serves as a computer-readable storage medium and can be used to store one or more programs. These programs can be software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the flight control optimization method provided in the embodiments of the present invention. Processor 710 executes the software programs, instructions, and modules stored in memory 720 to perform various functional applications and data processing in the electronic device, thereby implementing the flight control optimization method described in the aforementioned method embodiment.

[0173] The memory 720 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and applications required for at least one function; the data storage area may store data created based on the use of the electronic device, etc. In addition, the memory 720 may include a high-speed random access memory and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 720 may further include a memory remotely located relative to the processor 710, and these remote memories may be connected to the device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0174] The input device 730 may be used to receive input digital or character information and generate key signal input related to user settings and function control of the electronic device. The output device 740 may include a display device such as a display screen.

[0175] The present invention also provides a computer-readable storage medium storing a computer program executable by an electronic device. When the computer program runs on the electronic device, the electronic device executes the steps of a flight control optimization method.

[0176] Specifically, the computer storage medium of the embodiment of the present invention can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination of the above. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium can be any tangible medium that contains or stores a program, which can be used by or in combination with an instruction execution system, device or device.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A flight control optimization method, characterized in that: include: performing data normalization and integrity processing in response to the acquired flight status data, wherein the flight status data includes original time series data sets of pilot operation instructions, aircraft attitude parameters, and environmental monitoring indicators; Constructing an axial attention dynamic model to establish state transition relationships, including constructing an axial transformer network based on a dual-path attention mechanism to learn the spatiotemporal dependencies of flight state transitions. The axial attention dynamic model includes a time dimension attention layer and a feature dimension attention layer. The time dimension attention layer captures long-term dependency features in the operation sequence and generates a time dimension feature representation. In the feature dimension attention layer, the coupling relationship between different flight state data is learned to construct a multi-dimensional feature interaction matrix. Independently adjusted time dimension and feature dimension multi-head attention components are configured to enhance the representation capability of the axial attention dynamic model. A hierarchical network architecture containing residual connections is constructed to prevent degradation during the training process. Configure the deep reinforcement learning framework and define the multi-objective optimization mechanism, including the state space definition, action space mapping, and reward mechanism of the multi-objective optimization learning framework; In response to the processed flight status data, based on the axial attention dynamic model and the deep reinforcement learning framework, the network parameters are iteratively updated through an interactive training process to achieve gradient optimization of the flight control strategy, and the network parameters are initialized according to the differentiation strategy, including using an orthogonal matrix to initialize the policy network parameters and initializing the value evaluation network parameters based on the variance stability principle; a loss function based on the value evaluation network is constructed, including constructing a mean square error loss function of the time series difference error to optimize the prediction accuracy of the value evaluation network; an adaptive learning rate mechanism is set, including implementing a staged attenuation learning rate adjustment mechanism to balance training speed and convergence stability; a priority sampling strategy is constructed based on the sample importance weight to accelerate the learning process of key samples; In response to the flight status data after training, a dynamic association model of environmental factors and control parameters is output, a flight adjustment instruction set for deployment is generated, and a real-time influence coefficient matrix of environmental interference parameters and flight control parameters is calculated; based on the real-time influence coefficient matrix, a three-dimensional dynamic surface map reflecting the environment-operation association relationship is generated, the horizontal axis of the three-dimensional dynamic surface map maps the environmental parameters, the vertical axis represents the control parameters, and the surface height indicates the adjustment efficiency; the thrust compensation coefficient and the attitude conservative adjustment amount are dynamically output according to the risk level of the environmental parameters; the strategy robustness is tested in a flight simulation environment, and the trajectory tracking accuracy and energy consumption indicators are quantitatively evaluated.

2. A flight control optimization method according to claim 1, characterized in that: In response to the acquired flight status data, data standardization and integrity processing are performed, wherein the flight status data includes original time series data sets of pilot operation instructions, aircraft attitude parameters, and environmental monitoring indicators, further including: Obtaining the pilot's operating instructions, including obtaining physical quantities of the operating instructions, obtaining the aircraft's attitude parameters, including obtaining attitude angle parameters, and obtaining the environmental monitoring indicators, including obtaining atmospheric environment monitoring data; performing differential normalization processing on the flight status data, including performing a first type of range compression transformation on the pilot operation command data and a second type of symmetric range transformation on the aircraft attitude parameter data, wherein the first type of range compression transformation is a unit dimension interval and the second type of symmetric range transformation is a symmetric interval; In response to the flight status data after differential normalization, a time series interpolation algorithm is used to compensate for missing data segments, and outlier data points are identified and removed based on a probability distribution model; Construct a structured dataset, which involves splitting continuous time series data into training, validation, and test subsets with fixed time windows.

3. The flight control optimization method according to claim 1, characterized in that: Configure the deep reinforcement learning framework and define the multi-objective optimization mechanism, including the state space definition, action space mapping, and reward mechanism of the multi-objective optimization learning framework, further including: The state space definition includes constructing a multi-dimensional state vector of operational control variables, flight attitude variables, and environmental disturbance variables; The motion space mapping includes defining a continuous motion output space reflecting thrust adjustment instructions and attitude correction instructions; The reward mechanism includes constructing a penalty function based on the target state deviation measurement and introducing environmental safety boundary constraints; Deploy a random strategy generation network for action decision-making and simultaneously build a value evaluation network.

4. The flight control optimization method according to claim 1, characterized in that: The deep reinforcement learning framework also includes: A stabilization training mechanism for target network separation, including the use of dual-target network architecture separation strategy improvement and value estimation process; Integrate environmental over-limit penalty terms and safe operation reward terms into the basic reward function; Inject parameterized noise signals into the action selection layer to expand the strategy search space; Configure a discount factor close to unity to balance immediate operational benefits with long-term flight objectives.

5. A flight control optimization system, based on a flight control optimization method according to any one of claims 1 to 4, characterized in that: include: a data acquisition module configured to perform data normalization and integrity processing in response to acquired flight status data, wherein the flight status data includes original time series data sets of pilot operation instructions, aircraft attitude parameters, and environmental monitoring indicators; a dynamic model building module configured to build an axial attention dynamic model to establish a state transition relationship, including building an axial transformer network based on a dual-path attention mechanism for learning the spatiotemporal dependency of flight state transitions; A learning framework establishment module is configured to configure a deep reinforcement learning framework and define a multi-objective optimization mechanism, including a multi-objective optimization learning framework for state space definition, action space mapping, and reward mechanism; a data training module configured to iteratively update network parameters through an interactive training process in response to the processed flight state data based on the axial attention dynamic model and the deep reinforcement learning framework to achieve gradient optimization of the flight control strategy; The flight control optimization module is configured to output a dynamic association model between environmental factors and control parameters in response to the flight status data after training, and generate a flight adjustment instruction set for deployment.

6. An electronic device, characterized in that: include: A processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that It stores a computer program that can be executed by an electronic device. When the computer program runs on the electronic device, the electronic device executes the steps of the method according to any one of claims 1 to 4.

Images