A flight path stability optimization method, system, storage medium and device

Abstract

The present invention belongs to the field of path prediction and discloses a flight path stability optimization method, system, storage medium, and device. The method comprises constructing a dynamic fusion dataset by aligning timestamps in response to aircraft dynamics parameters, pilot physiological state parameters, and three-dimensional spatial environmental parameters acquired synchronously from multiple sources. A pre-built dual-channel feature extraction architecture is used to extract dual-channel features from the dynamic fusion dataset. The first channel uses a sparse attention mechanism to analyze dynamic coupling relationships in the time dimension, while the second channel uses a graph neural network to model the spatial correlation topology of environmental parameters. The dual-channel features are then integrated to construct a deep reinforcement learning decision engine, establishing a mapping strategy from environmental state to control action to path stability, and driving control strategy optimization through a reward mechanism. Finally, a real-time decision engine is used to generate multi-parameter collaborative optimization instructions for flight control surfaces. The present invention can improve the efficiency and accuracy of path prediction, thereby optimizing the path.

Description

Technical Field

[0001] The present invention belongs to the technical field of flight path prediction, and in particular relates to a flight path stability optimization method, system, storage medium and device. Background Art

[0002] The current mainstream flight path stability analysis methods have the following limitations:

[0003] Data mode simplification:

[0004] Existing systems primarily rely on aircraft dynamics parameters (such as pitch angle and airspeed) for decision-making, lacking comprehensive awareness of the pilot's physiological state (such as EEG fluctuations and heart rate variability) and high-dimensional environmental disturbances (such as wind shear and pressure gradients). This can lead to inaccurate control commands in the event of sudden turbulence or pilot stress, causing significant deviations from the flight path.

[0005] Long-term dependency modeling flaws:

[0006] Conventional LSTM / GRU models suffer from the vanishing gradient problem when processing minute-level flight data. While the standard Transformer architecture can capture long-range dependencies, its computational complexity makes it difficult to meet the demands of real-time control. For example, on a transoceanic flight, over an hour of flight path data can cause the computational load of a traditional attention mechanism to surge by over 300%.

[0007] Missing spatial association of the environment:

[0008] Existing methods treat environmental parameters like wind speed and temperature as independent variables, ignoring the fluid dynamics coupling between meteorological elements (such as the Navier-Stokes equations for wind speed and pressure). This simplified model can lead to path prediction errors exceeding 15% under complex meteorological conditions.

[0009] Therefore, the present application provides a flight path stability optimization method to solve the above technical problems. Summary of the Invention

[0010] The purpose of the present invention is to provide a flight path stability optimization method, system, storage medium and device to solve the technical problems in the prior art of large errors and low efficiency caused by single data modality and complex calculations during path prediction.

[0011] In order to solve the above technical problems, the present invention provides a flight path stability optimization method, comprising:

[0012] In response to the aircraft dynamic parameters, pilot physiological state parameters and three-dimensional space environment parameters collected synchronously by multiple sources, a dynamic fusion dataset is constructed by timestamp alignment;

[0013] In response to the dynamic fusion dataset, extracting dual-channel features of the dynamic fusion dataset based on a pre-built dual-channel feature extraction architecture, wherein the first channel analyzes the dynamic coupling relationship in the time dimension through a sparse attention mechanism, and the second channel models the spatial correlation topology of environmental parameters through a graph neural network;

[0014] The dual-channel features are integrated to build a deep reinforcement learning decision engine, establish a mapping strategy between environment state, control action and path stability, and drive control strategy optimization through a reward mechanism;

[0015] In response to the optimized control strategy, a multi-parameter collaborative optimization instruction of the flight control surface is generated based on a real-time decision engine to improve the path stability.

[0016] In some specific embodiments, in response to aircraft dynamic parameters, pilot physiological state parameters, and three-dimensional space environment parameters collected synchronously from multiple sources, constructing a dynamic fusion dataset by aligning timestamps further includes:

[0017] Collect the six-degree-of-freedom motion parameters of the flight control system, including thrust, three-axis rudder surface deflection roll angle, pitch angle, speed and heading angle, as aircraft dynamic parameters;

[0018] Collect pilot neurophysiological response parameters, including heart rate variability and EEG rhythm characteristics, as pilot physiological status parameters;

[0019] Collect multi-physical field parameters of the atmospheric environment, including wind speed vector, atmospheric pressure field, and temperature gradient field, as three-dimensional space environmental parameters;

[0020] A cross-modal data spatiotemporal alignment model is established to generate a fusion feature vector containing multi-dimensional parameters and dynamically fuse datasets.

[0021] In some specific embodiments, the first channel resolves the dynamic coupling relationship in the time dimension through a sparse attention mechanism, further comprising:

[0022] Embed multidimensional spatiotemporal dynamic parameters into multidimensional feature space;

[0023] Construct a local-global attention coupling mechanism to constrain the computational complexity of long-range dependencies through a configurable sparse window;

[0024] A cascaded encoder structure is used to extract features layer by layer. Each encoding unit contains a sparse attention sublayer and a fully connected feedforward sublayer.

[0025] Output is the feature tensor that characterizes the dynamic response of the aircraft-pilot system.

[0026] In some specific embodiments, the second channel models the spatial correlation topology of environmental parameters through a graph neural network, further comprising:

[0027] Map atmospheric environment parameters into dynamic graph nodes, and the node attributes include scalar parameters and vector parameters;

[0028] Constructing connection edges between nodes based on the fluid dynamics relationship between meteorological elements, wherein the connection edges include connection weights between nodes;

[0029] Perform multi-order graph convolution operations, where each order includes adjacency matrix regularization, degree matrix normalization, and feature nonlinear aggregation;

[0030] Output is a graph embedding vector that characterizes the propagation characteristics of environmental disturbances.

[0031] In some specific embodiments, the dual-channel features are integrated to construct a deep reinforcement learning decision engine, establish a mapping strategy between environment state, control action, and path stability, and drive control strategy optimization through a reward mechanism, further comprising:

[0032] Construct a multi-dimensional hybrid feature space as the environmental state representation, which includes the temporal dynamic features output by the first channel and the spatial topological features output by the second channel;

[0033] Define a four-dimensional continuous action space as the control action output, corresponding to the control increments of the power system, control surface, roll axis and pitch axis respectively;

[0034] Define a multi-objective reward function, including path deviation penalty, speed fluctuation penalty, and stability gain reward;

[0035] An actor-critic framework is adopted to train the policy network and achieve real-time closed-loop optimization of control actions.

[0036] In some specific embodiments, an actor-critic framework is used to train a policy network to achieve real-time closed-loop optimization of control actions, further comprising:

[0037] When using a deep Q-network, the target network is combined with temporal difference learning to achieve iterative updates of the value function;

[0038] When using proximal strategy optimization, the strategy update amplitude is constrained by importance sampling and gradient clipping;

[0039] Establish a causal relationship model between control increment parameters and path stability;

[0040] The contribution of attitude slope performance to path stability is verified through reward mechanism feedback.

[0041] In some specific embodiments, in response to the optimized control strategy, generating multi-parameter collaborative optimization instructions for flight control surfaces based on a real-time decision engine further includes:

[0042] Deploy a lightweight inference engine to process fused feature streams in real time;

[0043] Predict the combined impact of multi-control dimension adjustments on path stability;

[0044] Generate a collaborative optimization instruction set including thrust compensation, rudder correction, roll compensation angle, and pitch compensation angle;

[0045] A hierarchical warning and operation guidance mechanism is triggered based on the changing trend of the reward function.

[0046] Another aspect of the present invention provides a flight path stability optimization system, comprising:

[0047] a data acquisition module configured to construct a dynamic fusion data set by aligning timestamps in response to aircraft dynamic parameters, pilot physiological state parameters, and three-dimensional space environment parameters collected synchronously from multiple sources;

[0048] an architecture extraction module configured to extract dual-channel features of the dynamic fusion dataset based on a pre-built dual-channel feature extraction architecture in response to the dynamic fusion dataset, wherein the first channel analyzes the dynamic coupling relationship in the time dimension through a sparse attention mechanism, and the second channel models the spatial correlation topology of environmental parameters through a graph neural network;

[0049] a strategy optimization module configured to integrate the dual-channel features to construct a deep reinforcement learning decision engine, establish a mapping strategy between environment state, control action, and path stability, and drive control strategy optimization through a reward mechanism;

[0050] The decision suggestion module is configured to generate multi-parameter collaborative optimization instructions for the flight control surfaces based on the real-time decision engine in response to the optimized control strategy to improve the path stability.

[0051] A third aspect of the present invention provides an electronic device, comprising:

[0052] 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 executes the aforementioned flight path stability optimization method.

[0053] In a fourth aspect of the present invention, a computer-readable storage medium is proposed, which 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 above-mentioned flight path stability optimization method.

[0054] Compared with the prior art, the beneficial effects are:

[0055] The present invention discloses a flight path stability optimization method, system, storage medium and device. Through a multimodal data fusion architecture and a sparse attention-graph neural dual-channel feature extraction mechanism, combined with a deep reinforcement learning decision engine, it can effectively reduce the complexity of calculation, improve prediction efficiency and accuracy, and thus optimize the path. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of a flight path stability optimization method according to some specific embodiments of the present invention;

[0057] Figure 2 1. It is a schematic structural diagram of a flight path stability optimization system according to some specific embodiments of the present invention;

[0058] Figure 3 is a schematic structural diagram of an electronic device according to some specific embodiments of the present invention;

[0059] Reference numerals:

[0060] 710: processor; 720: memory; 730: input device; 740: output device. DETAILED DESCRIPTION

[0061] 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.

[0062] 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.

[0063] 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.

[0064] 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.

[0065] 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.

[0066] 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.

[0067] 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.

[0068] Reference Figure 1 , a flight path stability optimization method, comprising:

[0069] S101, in response to aircraft dynamics parameters, pilot physiological state parameters, and three-dimensional space environment parameters collected synchronously from multiple sources, constructing a dynamic fusion data set by aligning timestamps;

[0070] Specifically, in an embodiment of the present invention, a distributed sensor network is used to synchronously collect three types of flight parameters: aircraft dynamic parameters are obtained through the Honeywell HFR5-D flight data recorder, including thrust, rudder deflection angle, roll angle, pitch angle, speed and heading angle, a total of 6 parameters, and the sampling frequency is fixed at 64 times per second (64Hz). For example, during the cruise phase, the thrust value is recorded as 82.3kN and the roll angle is -1.2°; the pilot's physiological state parameters are captured using the BioNomadix multi-channel biological signal acquisition system, including two types of parameters: heart rate (ECG channel 500Hz sampling, such as basal heart rate 72bpm) and brain wave (EEG channel 1000Hz sampling, such as alpha wave amplitude 18.7μV); three-dimensional space environment parameters are captured by the Gill A WindMaster 3D ultrasonic anemometer (sampling wind speed and direction at 32Hz, such as monitoring a 25.7m / s crosswind) collaborates with a Bosch BMP388 sensor (sampling air pressure at 10Hz and temperature at 10Hz, such as recording an air pressure of 1013.25hPa and a temperature of -42.7°C) to collect data. The time synchronization system uses the Precision Time Protocol (PTP) for cross-device clock synchronization. Using GPS satellite timing signals as the reference time source, the synchronization error is controlled within ±0.5 milliseconds, and timestamps are uniformly formatted in UTC (e.g., 2025-07-09T08:00:00.000000000Z). Timestamp standardization includes: converting the flight data recorder's raw data packet count into absolute time. For example, when the data packet count is 1024 (15.625 milliseconds per packet), the timestamp is calculated as the base time plus 1024×15,625,000 nanoseconds, resulting in 2025-07-09T08:00:16.000000000Z; compensating for the 8-millisecond delay of the anemometer due to ultrasonic propagation, and resampling and interpolation processing using 1000Hz (1-millisecond interval) as the unified standard: flight control parameters (64Hz) use the cubic spline interpolation algorithm. For example, at 8:00:00.123Z, the roll angle is -0.5 based on the adjacent time point 8:00:00.120Z. The roll angle at 8:00:00.125Z is 1.2°. Interpolation is calculated using distance weighting: the time offset ratio is 0.6 (3 ms from the previous point / total interval is 5 ms). The interpolation result is (-0.5) × 0.4 + (1.2) × 0.6 = 0.42°. The air pressure data (10 Hz) uses linear interpolation. For example, if the air pressure is 1013.2 hPa at 8:00:00.100Z and 1013.0 hPa at 8:00:00.200Z, the interpolation at 8:00:00.123Z is: the time offset of 23 milliseconds accounts for 23% of the 100-millisecond interval. The interpolation result is 1013.2 - (1013.2 - 1013.0) × 0.23 = 1013.154 hPa.Time window alignment uses a 500-ms sliding window (100-ms step size). Kalman filtering prediction is used when the time deviation of the data within the window exceeds 1 ms. For example, if the EEG signal is missing at 8:00:05.350Z within the target time window 8:00:05.000-8:00:05.500Z, the predicted value is calculated based on the previous data points 8:00:05.348Z (12 μV) and 8:00:05.349Z (15 μV), using weighting factors of 0.6 (closest point), 0.3 (next closest point), and 0.1 (baseline value): 0.6 × 15 + 0.3 × 12 + 0.1 × 10 = 14.2 μV. Finally, a 12-dimensional fused feature vector is constructed, generating a complete data record every millisecond.

[0071] In some of these applications, in response to the aircraft dynamic parameters, pilot physiological state parameters and three-dimensional space environment parameters collected synchronously from multiple sources, a dynamic fusion data set is constructed through timestamp alignment, further including: collecting the six-degree-of-freedom motion parameters of the flight control system, including thrust, three-axis rudder surface deflection roll angle, pitch angle, speed and heading angle, as aircraft dynamic parameters; collecting the pilot's neurophysiological response parameters, including heart rate variability characteristics and EEG rhythm characteristics, as pilot physiological state parameters; collecting the atmospheric environment multi-physical field parameters, including wind speed vector, atmospheric pressure field, temperature gradient field, as three-dimensional space environment parameters; establishing a cross-modal data spatiotemporal alignment model to generate a fusion feature vector containing multi-dimensional parameters, and dynamically fuse the data set.

[0072] It can be understood that in this application, the six-degree-of-freedom motion parameter collection includes: net engine output thrust (unit: kN, example cruise value: 82.3kN); three-axis control surface deflection angle - aileron deflection angle controls the roll axis (range: ±25°, accuracy: 0.01°, example: 4.2° left deviation), rudder deflection angle controls the yaw axis (range: ±30°, accuracy: 0.01°, example: 1.78° right deviation), elevator deflection angle controls the pitch axis (range: ±35°, accuracy: 0.01°, example: 3.24° downward deviation); attitude angle - roll angle (range: ±180°, accuracy: 0.01°, example: -1.2°), pitch angle (range: ±90°, accuracy: 0.01°, example: 3.8°); motion state - speed (airspeed value, unit: m / s, example: 235m / s), heading angle (magnetic heading, unit: degree, example: 142.5°).

[0073] Neurophysiological response parameter processing: Heart rate variability (HRV) characteristics were used to extract the standard deviation of the RR interval (SDNN), measured in milliseconds. The calculation formula is: SDNN is equal to the sum of the squared deviations of the RR interval sequence from the mean divided by the square root of the number of RR intervals minus one, where the number of RR intervals, N, is taken from a 5-minute window (for example, SDNN = 42 milliseconds at a baseline heart rate of 72 beats / minute, rising to 15 milliseconds in a stressed state). EEG rhythm characteristics were divided into the power proportions of the delta wave (1-3 Hz), theta wave (4-7 Hz), and alpha wave (8-13 Hz) frequency bands (for example, the theta wave proportion rises to 52% in a fatigued state and 38% in an awake state).

[0074] The collection of multi-physics field parameters of the atmospheric environment includes: three-dimensional wind speed vector (for example, the east component U is 12.3m / s, the north component V is -8.7m / s, and the vertical component W is 0.3m / s); the atmospheric pressure field is characterized by the horizontal pressure gradient (for example, the nose-tail pressure difference is 0.8hPa, and the calculation formula is that the pressure gradient is equal to the tail pressure value minus the nose pressure value divided by the fuselage length of 60 meters); the temperature gradient field records the vertical temperature difference (for example, the gradient of 100 meters in altitude is -0.65℃). Measured data for the wind shear event (2025-07-09T09:31:05.500Z): wind speed vector [U=25.7m / s, V=-12.3m / s, W=-4.2m / s], pressure field distribution (nose 1012.4 hPa / left wing 1012.8 hPa / right wing 1012.1 hPa), temperature gradient (-42.7°C at the belly and -43.2°C at the back, forming a vertical gradient of -0.5°C / 10 meters).

[0075] The cross-modal spatiotemporal alignment model is based on the UTC atomic clock time reference (synchronization error <0.1 millisecond) and spatial registration to the aircraft coordinate system (origin at the aircraft's center of gravity). The fusion process includes: timestamp normalization (e.g., compensating for 0.005-second delay in EEG devices); mapping environmental parameters to the aircraft coordinate system via the rotation matrix R (constructed from real-time roll / pitch / yaw angles); and generating a 36-dimensional dynamic tensor: 8 dimensions for six-degree-of-freedom parameters (thrust / rudder / attitude, etc.), 6 dimensions for neurophysiological parameters (heart rate / SDNN / δθαβ wave power ratio), and 22 dimensions for the environmental physical field (5-point 3D wind speed, 5-point pressure, 5-point temperature, and 3 gradients). This multimodal fusion architecture provides high-fidelity input for sparse attention feature extraction and reinforcement learning decision-making.

[0076] S102, in response to the dynamic fusion dataset, extracting dual-channel features of the dynamic fusion dataset based on a pre-built dual-channel feature extraction architecture, wherein the first channel analyzes the dynamic coupling relationship in the time dimension through a sparse attention mechanism, and the second channel models the spatial correlation topology of environmental parameters through a graph neural network;

[0077] Specifically, in this embodiment of the present invention, a dual-channel parallel architecture is used to process the fused dataset: the first channel is used to analyze the time series characteristics of aircraft dynamic parameters and pilot physiological parameters, while the second channel processes the spatial topological relationships of flight environment parameters. The input data is a 36-dimensional fused dataset (with a temporal resolution of 1 millisecond).

[0078] Time feature channel implementation: 14 time series parameters (8-dimensional flight parameters + 6-dimensional physiological parameters) are input, with a data point per millisecond, and 60,000 time steps (corresponding to 1 minute of flight data) are processed continuously. For example, the input segment (timestamp 2025-07-09T08:10:23.500Z) contains: thrust 81.4 kN, aileron deflection angle 3.7 degrees, roll angle 3.8 degrees, pitch angle 4.1 degrees, speed 226 m / s, heading angle 142.5 degrees, heart rate 89 beats / minute, and theta wave power percentage 53%). The processing flow includes:

[0079] High-dimensional embedding conversion: The input is mapped to a 256-dimensional feature space using a 14×256-dimensional embedding matrix. For example, the thrust parameter 81.4 kN is converted to the feature vector [0.32, -1.45, ..., 0.87].

[0080] Local attention window calculation: A 64-step sliding window (covering 6.4 seconds of data) is set, and only the association between the current time point and the data points within 32 steps before and after is calculated. For example, at 2025-07-09T08:10:23.506Z, the association weight with the time 32 steps ago (2025-07-09T08:10:23.474Z) is calculated to be 0.83. Data points outside the window range (such as 2025-07-09T08:10:17.106Z, 640 steps ago) are blocked.

[0081] Cascade encoding: A three-layer encoder architecture is used, with each layer comprising an eight-head sparse attention module and a 512-node feedforward neural network. Residual connections are used to enhance features (input + processing results). For example, a roll angle of 3.8 degrees outputs [0.12,...] in the first layer, which is enhanced to [1.32,...] in the third layer. The final output is a 256-dimensional temporal feature vector. Key features include dimension 38 (the coupling strength between roll angle mutations and theta surges, with an example value of 1.32) and dimension 112 (the correlation between speed descent rate and heart rate variability, with an example value of 0.94).

[0082] Spatial feature channel implementation: Construct an environment graph structure containing 15 nodes, whose attributes include parameters such as the three-dimensional wind speed component at the left wing tip, nose air pressure, vertical temperature gradient, etc. Edge weights are calculated based on physical relationships:

[0083] The wind speed-pressure edge weight is 1 divided by 1 plus the square of the pressure difference (e.g., a pressure difference of 0.8 hPa has a weight of 0.61).

[0084] The weight of the temperature-wind speed edge is equal to the negative temperature difference of e divided by 10 (for example, a temperature difference of 0.5 degrees Celsius has a weight of 0.95);

[0085] The graph convolution operation is processed in three layers:

[0086] Adjacency matrix optimization: adding self-loops (node ​​self-connection weight 1.0), degree matrix normalization;

[0087] Feature update: Each layer performs a nonlinear transformation, and the formula is that the new feature is equal to the result of the ReLU activation function applied to the normalized adjacency matrix multiplied by the input feature multiplied by the weight matrix;

[0088] Dimensionality transformation: First layer: 15-D input → 32-D output (for example, nose pressure node [1012.4, 0.015] → [0.32, 1.15, -0.83, ...]), second layer: 32-D → 64-D, and third layer: 64-D → 128-D output (for example, [-0.23, 0.76, ..., 1.05]). Key features include dimension 17 (vertical wind shear intensity, example value 1.05) and dimension 89 (temperature-wind speed coupling coefficient, example value 0.63).

[0089] In some of these applications, the first channel uses a sparse attention mechanism to analyze the dynamic coupling relationship in the time dimension, including embedding multi-dimensional spatiotemporal dynamic parameters into a multi-dimensional feature space; constructing a local-global attention coupling mechanism to constrain the computational complexity of long-range dependencies through a configurable sparse window; using a cascaded encoder structure to extract features layer by layer, with each encoding unit containing a sparse attention sublayer and a fully connected feedforward sublayer; and outputting a feature tensor that represents the dynamic response of the aircraft-pilot system.

[0090] As you can understand, in this application, in response to the fused aircraft-pilot multimodal time series data, taking the eight-dimensional parameters (thrust, control surface deflection, roll angle, pitch angle, speed, heading angle, heart rate, and brain waves) of a Boeing 787 during cruise phase as an example, the sampling rate is 1000Hz. First, the original parameters are embedded in a 256-dimensional latent space: the thrust of 82.3kN is mapped to the vector [0.12, -0.24, ..., 0.18] (dimension 256) through a linear transformation layer, and the brain wave of 18.7μV is mapped to [-0.05, 0.33, ..., -0.11], with the embedding weight matrix size being 8×256. Then, a local-global attention coupling mechanism is constructed: a configurable sparse window is set to 200 milliseconds (covering 200 sampling points). At t=08:10:23.456Z, the attention weights of target position 456 and adjacent positions are calculated. Position 300 (156ms from the target) is blocked because it is outside the window, position 400 (56ms from the target) receives a weight of 0.15, and position 450 (6ms from the target) receives a weight of 0.72. At the same time, four global key positions are retained (including the takeoff time t=0 and the turbulence starting point t=23.500Z). The global position 23.500Z is associated with the current point with a weight of 0.28. A three-layer cascaded encoder extracts features layer by layer: In the first encoding unit, a sparse attention sublayer computes a weighted sum of positions within a 200-window window (for example, the cross-attention weight between a roll angle feature of -1.2° and a rudder deflection of 2.15° is 0.63), outputting a 128-dimensional vector. The feedforward sublayer processes this information using a fully connected network (128 → 256 → 128 neurons) and a ReLU activation function to generate the features [0.18, -0.32, ..., 0.09]. The second encoder layer inputs this feature and computes high-order coupling relationships within the same window (for example, the correlation weight between a sudden increase in EEG theta wave power and a pitch angle fluctuation three seconds later is 0.41). Finally, the third layer outputs a 24-dimensional dynamic response feature tensor, including key coupling metrics: a rudder-roll response coefficient of 0.87 (indicating a 0.1° rudder deflection causes a 0.23° roll angle change), and an EEG-thrust delayed correlation strength of 0.62 (a 10% increase in theta wave power leads to a 1.2% thrust increase two seconds later). Verification shows that during a 30-second turbulence event, this mechanism captured the millisecond-level dynamic coupling between sudden wind speed changes and the pilot's stress response (a 17 bpm increase in heart rate). The feature tensor's contribution to the flight path prediction task reached 0.79, a 46% increase compared to the traditional LSTM feature (weight of 0.54).

[0091] In another application, the second channel models the spatial correlation topology of environmental parameters through graph neural networks, including mapping atmospheric environmental parameters into dynamic graph nodes, with node attributes including scalar parameters and vector parameters; constructing connecting edges between nodes based on the fluid dynamics relationship between meteorological elements, and the connecting edges include connection weights between nodes; performing multi-order graph convolution operations, with each order operation including adjacency matrix regularization, degree matrix normalization and feature nonlinear aggregation; and outputting a graph embedding vector that characterizes the propagation characteristics of environmental disturbances.

[0092] It can be understood that in this application, in response to the multi-physics field parameters of the atmospheric environment, a dynamic graph structure is constructed, taking the wind speed vector, pressure field, and temperature gradient field collected during the cruise phase of a Boeing 787 (altitude FL350) as an example: the graph nodes are defined as five spatially distributed sensor locations (nose, left wingtip, right wingtip, belly, and vertical tail), and the node attributes include scalar parameters (such as nose temperature -42.7°C, left wingtip pressure 1012.8hPa) and vector parameters (such as the wind speed vector at the belly [12.3m / s, -8.7m / s, 0.3m / s]). Weighted connections are constructed based on meteorological fluid dynamics: the distance between the nose and the left wingtip node is 15 meters. The pressure-wind speed coupling weight is calculated according to the Navier-Stokes equations. For a pressure gradient difference of 0.5 hPa, the weight is 0.5 × exp(-0.2 × distance) (for example, an actual pressure difference of 0.8 hPa results in a weight of 0.38). The temperature-wind speed edge weight is calculated according to the Richardson number criterion. For a temperature gradient of -0.6°C / 100m, the weight is 0.7 × |ΔT| / 10 (for example, a temperature difference of 2.1°C between the vertical tail and the belly of the aircraft results in a weight of 0.15). A third-order graph convolution operation is performed: during the first-order aggregation of neighborhood features, self-loops are added to the adjacency matrix (the weight of the connection between node i and itself is set to 1.0). The degree matrix is ​​calculated as the inverse of the square root of the number of node connections (for example, the nose node degree is 3, which is normalized by a factor of 1 / √3≈0.58). Feature aggregation uses the linear transformation matrix W. （1） (size 8×16) and ReLU activation function, outputting the left wingtip node features [-0.12, 0.25, ..., -0.08] (16 dimensions). The second-order convolution is extended to the second-order neighborhood, and the nose node aggregates the first-order features of the left wingtip [-0.12, 0.25] and the second-order features of the right wingtip [0.18, -0.31] (weighted by degree matrix normalization). The right wingtip feature weight = 1 / √4 = 0.5, and is passed through the matrix W （2） After transformation (size 16×12), the features [0.15, -0.22, ..., 0.09] are generated. The third-order convolution introduces the global disturbance propagation characteristics to calculate the pressure wave transmission path (for example, when the nose pressure suddenly changes from 1013.5 to 1012.7 hPa, the propagation path with the highest weight is nose → left wing tip → belly, with a path weight of 0.38×0.42=0.16). （3）The aggregation output is an 8-dimensional graph embedding vector (size 12×8), where the wind shear propagation coefficient at index 3 is 0.73 (indicating that every 1 m / s change in vertical wind speed causes a 0.73 hPa / km change in pressure gradient). Validation: During a measured wind shear event (2025-07-09T11:23:45Z), this vector accurately predicted that a vertical gust would propagate from the belly of the aircraft to the vertical tail within 0.8 seconds (actual propagation time 0.82 seconds), with an environmental disturbance modeling error of less than 3%, an 86% improvement over the traditional independent parameter model (error of 21%).

[0093] S103, integrating the dual-channel features to build a deep reinforcement learning decision engine, establish a mapping strategy between environment state, control action, and path stability, and drive control strategy optimization through a reward mechanism;

[0094] Specifically, in this embodiment of the present invention, the temporal dynamic features (24 dimensions) extracted by the sparse attention architecture and the spatial topological features (8 dimensions) extracted by the graph neural network are concatenated into a 32-dimensional mixed feature vector as the environmental state input. Taking the approach phase of a Boeing 787 as an example, the decision engine construction process is as follows: First, the state space is defined as a mixed feature vector. The example state vector includes parameters such as the roll angle eigenvalue of -1.2 (normalized value), the EEG theta wave power ratio of 0.53, and the vertical wind speed gradient of -0.015. The action space is designed as a four-dimensional continuous vector: the first dimension is the thrust adjustment range of ±10% (the example value of +2.3% corresponds to an actual thrust increase of 5.6 kN), the second dimension is the rudder angle adjustment range of ±0.5° (example -0.12°), the third dimension is the roll angle adjustment range of ±0.3° (example +0.08°), and the fourth dimension is the pitch angle adjustment range of ±0.4° (example -0.15°). The reward function uses a weighted multi-objective calculation: the path deviation term is the lateral distance between the actual path and the planned route (for example, a deviation of 38 meters is converted to a penalty of -3.8); the speed fluctuation term is the absolute value of the airspeed change rate (for example, 0.8 m / s² is converted to -1.2); and the stability gain term is the attitude angle smoothness (for example, a roll angle standard deviation of 0.05° is converted to +2.1). The weight coefficients α = 1.0, β = 0.8, and γ = 1.2, resulting in a final reward of R = -3.8 × 1.0 - 1.2 × 0.8 + 2.1 × 1.2 = -2.16. The training phase uses a proximal policy optimization algorithm. The policy network has a three-layer fully connected structure (with 128 neurons, 64 neurons, and 32 neurons), a learning rate of 0.00025, and an experience replay buffer capacity of 10,000. Specific training iteration process: In a simulated wind shear scenario, the initial action output was [thrust +1.8%, rudder -0.10°, roll +0.12°, pitch -0.18°]. After execution, the path deviation increased to 42 meters, causing the reward to drop to -4.2. After 30,000 iterative optimizations, the action was optimized to [thrust +2.1%, rudder -0.09°, roll +0.05°, pitch -0.12°]. The deviation was reduced to 22 meters, causing the reward to increase to -0.8. A lightweight inference engine (model size 48MB) was deployed in the real-time decision-making phase. During a measured turbulence event (2025-07-09T11:23:45Z), the input detected a vertical gust load of +0.3G. The engine generated action commands within 8 milliseconds (thrust +3.2%, rudder +0.07°, roll -0.11°, pitch +0.06°), keeping the path deviation within 15 meters (reward function value -0.3), a 71% improvement over traditional PID control (deviation of 52 meters). Verification of the performance showed that after 2,000 simulated flight training flights, the heading stability index (path angle standard deviation) was reduced from the original 1.8° to 0.4°, and fuel efficiency was improved by 4.7%.

[0095] In some of these applications, the dual-channel features are integrated to construct a deep reinforcement learning decision engine, establish a mapping strategy between environment state, control action and path stability, and drive control strategy optimization through a reward mechanism, including constructing a multi-dimensional hybrid feature space as an environment state representation, including the temporal dynamic features output by the first channel and the spatial topological features output by the second channel; defining a four-dimensional continuous action space as the control action output, corresponding to the control increments of the power system, control surface, roll axis and pitch axis respectively; defining a multi-objective reward function, including a path deviation penalty term, a speed fluctuation penalty term and a stability gain reward term; and adopting an actor-critic framework to train the policy network to achieve real-time closed-loop optimization of the control action.

[0096] As you can understand, in this application, the 24-dimensional temporal dynamic features output by the sparse attention architecture (including parameters such as the control surface-roll response coefficient of 0.87 and the EEG-thrust delay correlation strength of 0.62) and the 8-dimensional spatial topological features generated by the graph neural network (including parameters such as the wind shear propagation coefficient of 0.73 and the pressure gradient correlation factor of -0.15) are concatenated into a 32-dimensional mixed feature vector as the environmental state input. Taking the Boeing 787 approach phase (at an altitude of 3000 meters) as an example, the state vector S(t) = [0.12, -0.82, 0.05, ..., 0.61, -0.27] is constructed (where indices 0-23 are temporal features and indices 24-31 are spatial features, with the example value -0.82 representing the normalized roll angle deviation). Define a four-dimensional continuous action space: the first-dimensional power system control increment corresponds to a thrust adjustment range of ±10% (for example, a +2.3% action is mapped to an actual thrust increase of 5.6 kN), the second-dimensional control surface control increment corresponds to a rudder angle adjustment range of ±0.5° (for example, -0.12°), the third-dimensional roll axis control increment range is ±0.3° (for example, +0.08°), and the fourth-dimensional pitch axis control increment range is ±0.4° (for example, -0.15°). Design a multi-objective reward function: The path deviation penalty is the lateral distance (in meters) between the actual path and the planned route, divided by 10 and negated (for example, a deviation of 38 meters is converted to -3.8). The speed fluctuation penalty is the absolute value of the airspeed change rate multiplied by a coefficient of 1.5 and negated (for example, a change rate of 0.8 m / s² corresponds to -1.2). The stability gain reward is the inverse of the attitude angle change smoothness multiplied by a coefficient of 0.1 (for example, a roll angle standard deviation of 0.05° corresponds to a reward of 2.0). The weights of the three components are α=1.0, β=0.8, and γ=1.2, respectively. The final reward R=(-3.8×1.0)+(-1.2×0.8)+(2.0×1.2)=-2.36. The policy network is trained using an actor-critic framework: the actor network is a three-layer, fully connected structure (128-64-32 neurons) that outputs action probability distributions; the critic network (128-64-1 structure) evaluates state values. A proximal policy optimization algorithm is used with a learning rate of 0.00025 and a replay buffer of 10,000 experience entries. An example of the training process: the initial policy outputs actions [+1.8%, -0.10°, +0.12°, -0.18°] in a wind shear scenario, resulting in a path deviation of 42 meters (a reward of -4.2). After 30,000 iterations of optimization, the policy evolves to [+2.1%, -0.09°, +0.05°, -0.12°], reducing the deviation to 22 meters (a reward of -0.8). During the real-time closed-loop optimization phase, when a vertical gust was detected (the wind shear coefficient in the eigenvector suddenly increased by 0.15), the decision engine generated an optimization action [+3.2%, +0.07°, -0.11°, +0.06°] within 8 milliseconds. After execution, the path deviation was controlled within 15 meters (a reward of -0.3).

[0097] Furthermore, an actor-critic framework is adopted to train the policy network to achieve real-time closed-loop optimization of control actions. This includes combining the target network with temporal difference learning to iteratively update the value function when using a deep Q-network; constraining the policy update amplitude through importance sampling and gradient clipping when using proximal policy optimization; establishing a causal association model between control increment parameters and path stability; and verifying the contribution of attitude slope performance to path stability through feedback from a reward mechanism.

[0098] It is understandable that the actor-critic framework training is implemented in the deep reinforcement learning decision engine, taking the cruise phase of the Boeing 787 as an example: when the deep Q network (DQN) is adopted, a dual network structure is deployed (online network parameters θ, target network parameters θ⁻), the online network outputs the action value Q(s,a;θ), and the target network calculates the temporal difference target: target value = immediate reward + discount factor 0.99 × the maximum Q value of the next state output by the target network. Example: Executing action a (thrust +2.3%) at state s at t=08:10:23.456Z yields a reward r=-2.36. The target network calculates the maximum Q-value for the next state s' to be 8.7, resulting in a target value of -2.36 + 0.99 × 8.7 = 6.23. The original Q-value estimated by the online network is 5.8, and is updated at a learning rate of 0.001: the new Q-value = 5.8 + 0.001 × (6.23 - 5.8) = 5.80043. The target network synchronizes θ⁻ = θ every 1000 steps. When using proximal policy optimization (PPO), the actor network outputs an action probability distribution π(a|s;θ). Importance sampling constrains the update amplitude: the probability ratio r(θ) is calculated as new policy probability / old policy probability, clipped to the interval [0.8, 1.2]. Example: The old policy outputs the action "roll +0.08°" with probability π_old = 0.25 in state s, and the new policy has probability π_new = 0.31. The probability ratio r = 0.31 / 0.25 = 1.24, which is clipped to 1.2. The loss function L = min(r(θ)A, clip(r(θ), 0.8, 1.2)A), where the advantage function A = actual reward minus the baseline value (for example: actual reward 6.5, critic network baseline 6.2 → A = 0.3), and the gradient update amplitude is limited to ±0.02. A causal relationship model between control increment and path stability is established: Regression analysis shows that a 1% increase in thrust adjustment ΔT reduces path deviation by 1.2 meters (regression coefficient β_T = -1.2, p < 0.01), and a 0.1° increase in roll angle adjustment Δφ reduces deviation by 0.8 meters (β_φ = -8.0). The contribution of attitude slope performance was verified through reward mechanism decomposition: after the roll angle adjustment action was executed, the contribution value of the stability gain term in the reward function increased to +2.5 (originally +1.8), and its proportion in the total reward increased from 32% to 41%.

[0099] S104 , in response to the optimized control strategy, generating multi-parameter collaborative optimization instructions for the flight control surface based on a real-time decision engine to improve path stability.

[0100] Specifically, in an embodiment of the present invention, in response to the optimized action vector (dimension 4) output by the deep reinforcement learning policy network, a multi-parameter coordinated instruction for the flight control surface is generated by an instruction mapping engine. For example, an Airbus A350 encountering a crosswind during approach (at an altitude of 500 meters) uses the decision engine to output a real-time action vector a = [+3.2%, -0.08°, +0.15°, -0.12°]. The command mapping engine converts the first-dimensional thrust adjustment (+3.2%) into an actual thrust command: the current thrust value is 142.6 kN × (1 + 3.2 / 100) = 147.1 kN. The second-dimensional rudder angle adjustment (-0.08°) is mapped into a rudder command: the base deflection angle of 2.1° is corrected to 2.02°. The third-dimensional roll angle adjustment (+0.15°) is converted into an aileron control variable: the left aileron is deflected downward by 0.10° and the right aileron is deflected upward by 0.05° (differential ratio 2:1). The fourth-dimensional pitch angle adjustment (-0.12°) corresponds to an elevator command: the current -1.8° is adjusted to -1.92°. The collaborative optimization logic is as follows: when the roll angle command exceeds a threshold of ±0.1°, rudder compensation is automatically coupled (with a proportional coefficient of 0.6). In this example, a roll command of +0.15° triggers an additional rudder compensation of +0.15×0.6=+0.09°, adjusting the final rudder command from 2.02° to 2.11°. The command generation latency is kept within 5 milliseconds. Effectiveness: After executing this collaborative command, the lateral path deviation caused by crosswinds was reduced from 32 meters before optimization to 7 meters, the standard deviation of the roll angle fluctuation improved from 0.25° to 0.07°, and the reward function value increased from -4.1 to +0.8.

[0101] In some of these applications, in response to the optimized control strategy, multi-parameter collaborative optimization instructions for the flight control surfaces are generated based on a real-time decision engine, including deploying a lightweight inference engine to process the fused feature flow in real time; predicting the comprehensive impact of multi-control dimension adjustments on path stability; generating a collaborative optimization instruction set including thrust compensation, rudder correction, roll compensation angle and pitch compensation angle; and triggering a graded warning and operation guidance mechanism based on the changing trend of the reward function.

[0102] It is understandable that in this application, a lightweight inference engine is deployed to process 32-dimensional fused feature streams (sampling rate 100Hz) in real time. Take the Airbus A350 approach phase at Frankfurt Airport (50.1°N, altitude 300 meters) as an example: when it is detected that the wind shear propagation coefficient in the feature vector suddenly increases to 0.85 (threshold 0.7) and the standard deviation of the roll angle fluctuation rises to 0.18° (threshold 0.1°), the engine predicts the combined impact of multi-control dimension adjustments within 5 milliseconds - thrust compensation of +3.2% can increase airspeed by 1.5m / s but increase fuel consumption by 2.1%, and roll compensation of -0.11° can reduce the path deviation by 12 meters but may cause the passenger's physical discomfort to increase by 0.3G. Based on the Pareto optimization criterion, a coordinated command set was generated: thrust compensation of +2.8% (a compromise solution, corresponding to an actual thrust increase from 142.6 kN to 146.5 kN); rudder correction of 0.07° left (from a base value of 2.1° to 2.17°); roll compensation of 0.08° left wing down (achieved through aileron differential: left aileron -0.12°, right aileron +0.04°); and pitch compensation of -0.09° (elevator adjusted from -1.8° to -1.89°). The coordinated logic implements coupled roll-pitch compensation: pitch compensation is automatically triggered when the absolute value of the roll command exceeds 0.05°. The calculation formula is: pitch compensation increment = -0.6 × roll compensation (for example, -0.6 × 0.08° = -0.048°, which, when added to the base value of -0.09°, results in a final value of -0.138°). Real-time monitoring of reward function trends: If the initial reward value is -1.8 and it does not rise to -1.0 within 3 seconds after the command is executed (the reward growth per minute is less than 0.4), a level 1 alarm is triggered. If there is no improvement for 6 seconds (the cumulative growth is less than 0.2), it is upgraded to a level 2 alarm and a fault code is automatically sent to the ground control console.

[0103] In another embodiment of the present invention, the following parameters are collected by a flight data recorder (FDR):

[0104] thrust , rudder angle , roll angle , pitch angle , speed , heading angle ;

[0105] Pilot physiological data: The following parameters are collected through wearable devices:

[0106] Heart rate , brain waves ;

[0107] Flight environment data: The following parameters are collected through meteorological sensors:

[0108] wind speed ,wind direction , air pressure ,temperature ;

[0109] The three types of data are time-aligned and fused to form a comprehensive data set. Each data record can be represented as:

[0110] Use sparse attention architecture to process time series features of flight data and pilot physiological data.

[0111] Sparse Attention Architecture: Input Embedding embeds time series data into a high-dimensional space. Sparse Attention Mechanism reduces computational complexity by sparsifying the attention matrix. For example, local attention or sparse attention are used. The encoder consists of multiple layers of sparse attention encoders, each layer including a sparse attention module and a feed-forward neural network. Output feature representation: After processing by the sparse attention encoder, the time series feature representation is obtained. .

[0112] Mathematical formula: Input embedding:

[0113]

[0114] .

[0115] Sparse Attention Mechanism:

[0116]

[0117] in, is the query, key, and value matrix, is the dimension of the key vector, is the sparse window size, used to limit the attention range.

[0118] Encoder output:

[0119]

[0120] A graph neural network (GNN) is used to process the spatial features of flight environment data. The flight environment data is modeled as a graph structure, where nodes represent different environmental parameters and edges represent the relationships between parameters.

[0121] Graph structure representation: Node features: ,Edge features define the weight of edges based on physical relationships (such as the relationship between wind speed and air pressure).

[0122] GNN architecture: Graph Convolution:

[0123]

[0124] in, is the adjacency matrix plus self-loops, is the degree matrix, is the weight matrix, is the activation function.

[0125] Mathematical formula: Initialize node features:

[0126]

[0127] -Graph convolution operation:

[0128]

[0129] - Output space feature representation:

[0130]

[0131] The time series features and spatial characteristics Fusion is performed to form a comprehensive feature representation:

[0132]

[0133] Use the deep reinforcement learning (DRL) framework, combined with the deep Q-network (DQN) or proximal policy optimization (PPO) algorithm, to learn the relationship between pilot operation and path stability.

[0134] Status definition:

[0135]

[0136] Action definition:

[0137]

[0138] in: Indicates the amount of thrust adjustment. Indicates the adjustment amount of the rudder angle. Indicates the adjustment amount of the roll angle. Indicates the adjustment amount of the pitch angle.

[0139] Reward function:

[0140]

[0141] in: Indicates path deviation. Indicates speed change. Indicates path stability. is the weight coefficient.

[0142] DRL model training: Use DQN or PPO algorithm to train the model and learn the optimal strategy , to maximize the cumulative reward.

[0143] Mathematical formula: Q function update (DQN):

[0144]

[0145]

[0146] Policy Update (PPO):

[0147]

[0148] The fused features The input is fed into the trained DRL model to predict in real time the impact of the pilot’s operation on the path stability.

[0149] Based on the model's prediction results, optimization suggestions are generated, such as adjusting thrust, changing rudder deflection, correcting roll angle and pitch angle, etc., to help pilots adjust operations in real time and improve path stability.

[0150] For example, at a certain moment , the pilot adjusted the thrust and rudder angle , while adjusting the roll angle and pitch angle The model evaluates the effect of these adjustments on path stability by performing the following steps:

[0151] The adjusted input energy and posture slope energy are input as time series features into the sparse attention architecture to obtain feature representation .

[0152] The time series features and spatial characteristics Fusion to form a comprehensive feature representation .

[0153] The comprehensive feature representation Input into the DRL model, the model is based on the current state and actions , calculate the reward function .

[0154] If the adjusted path deviation Decrease, speed change Maintain stability and path stability Improve, then the reward function The value of will increase, indicating that the adjustments are effective.

[0155] If the adjusted path deviation Increase, speed change Violent and stable path Decreases, then the reward function The value of will decrease, indicating that these adjustments are ineffective.

[0156] 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.

[0157] like Figure 2 As shown, a flight path stability optimization system according to a second embodiment of the present invention includes: a data acquisition module configured to construct a dynamic fusion data set by aligning timestamps in response to aircraft dynamic parameters, pilot physiological state parameters, and three-dimensional space environment parameters acquired synchronously from multiple sources;

[0158] an architecture extraction module configured to extract dual-channel features of the dynamic fusion dataset based on a pre-built dual-channel feature extraction architecture in response to the dynamic fusion dataset, wherein the first channel analyzes the dynamic coupling relationship in the time dimension through a sparse attention mechanism, and the second channel models the spatial correlation topology of environmental parameters through a graph neural network;

[0159] a strategy optimization module configured to integrate the dual-channel features to construct a deep reinforcement learning decision engine, establish a mapping strategy between environment state, control action, and path stability, and drive control strategy optimization through a reward mechanism;

[0160] The decision suggestion module is configured to generate multi-parameter collaborative optimization instructions for the flight control surfaces based on the real-time decision engine in response to the optimized control strategy to improve the path stability.

[0161] 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.

[0162] It should be noted that the aforementioned embodiment provides a flight path stability optimization system, using the division of the aforementioned functional modules as an example. In actual applications, the aforementioned 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 aforementioned embodiments can be combined into a single 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 considered as undue limitations of the present invention.

[0163] like Figure 3 As shown, an electronic device according to a third embodiment of the present invention includes:

[0164] 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 executes the aforementioned flight path stability optimization method.

[0165] A computer-readable storage medium according to a fourth embodiment of the present invention stores a computer program executable by an electronic device. When the computer program runs on the electronic device, the electronic device executes the above-mentioned flight path stability optimization method.

[0166] 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 memory and processor described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0167] Figure 3 Schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. Figure 3 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 3 A processor 710 is taken 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 path stability optimization method as described in any one of the embodiments of the present invention.

[0168] The electronic device may further include an input device 730 and an output device 740 .

[0169] 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 3 The bus connection is taken as an example.

[0170] 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 path stability 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 path stability optimization method described in the aforementioned method embodiment.

[0171] 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.

[0172] 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.

[0173] The computer storage medium of the embodiment of the present invention may adopt any combination of one or more computer-readable media. The computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media 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 thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device.

[0174] 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 path stability optimization method, characterized in that: include: In response to the aircraft dynamic parameters, pilot physiological state parameters and three-dimensional space environment parameters collected synchronously by multiple sources, a dynamic fusion dataset is constructed by timestamp alignment; In response to the dynamic fusion dataset, extracting dual-channel features of the dynamic fusion dataset based on a pre-built dual-channel feature extraction architecture, wherein the first channel analyzes the dynamic coupling relationship in the time dimension through a sparse attention mechanism, and the second channel models the spatial correlation topology of environmental parameters through a graph neural network; The dual-channel features are integrated to build a deep reinforcement learning decision engine, establish a mapping strategy between environment state, control action, and path stability, drive control strategy optimization through a reward mechanism, and construct a multi-dimensional hybrid feature space as an environmental state representation, which includes the temporal dynamic features output by the first channel and the spatial topological features output by the second channel. Define a four-dimensional continuous action space as the control action output, corresponding to the control increments of the power system, control surface, roll axis and pitch axis respectively; Define a multi-objective reward function, including path deviation penalty, speed fluctuation penalty, and stability gain reward; Adopting an actor-critic framework to train the policy network and achieve real-time closed-loop optimization of control actions; In response to the optimized control strategy, a multi-parameter collaborative optimization instruction of the flight control surface is generated based on a real-time decision engine to improve the path stability.

2. The flight path stability optimization method according to claim 1, characterized in that: In response to aircraft dynamic parameters, pilot physiological state parameters, and three-dimensional space environment parameters collected synchronously from multiple sources, a dynamic fusion data set is constructed by timestamp alignment, further comprising: Collect the six-degree-of-freedom motion parameters of the flight control system, including thrust, three-axis rudder surface deflection roll angle, pitch angle, speed and heading angle, as aircraft dynamic parameters; Collect pilot neurophysiological response parameters, including heart rate variability and EEG rhythm characteristics, as pilot physiological status parameters; Collect multi-physical field parameters of the atmospheric environment, including wind speed vector, atmospheric pressure field, and temperature gradient field, as three-dimensional space environmental parameters; A cross-modal data spatiotemporal alignment model is established to generate a fusion feature vector containing multi-dimensional parameters and dynamically fuse datasets.

3. The flight path stability optimization method according to claim 1, characterized in that: The first channel analyzes the dynamic coupling relationship in the time dimension through a sparse attention mechanism, further comprising: Embed multidimensional spatiotemporal dynamic parameters into multidimensional feature space; Construct a local-global attention coupling mechanism to constrain the computational complexity of long-range dependencies through a configurable sparse window; A cascaded encoder structure is used to extract features layer by layer. Each encoding unit contains a sparse attention sublayer and a fully connected feedforward sublayer. Output is the feature tensor that characterizes the dynamic response of the aircraft-pilot system.

4. The flight path stability optimization method according to claim 1, characterized in that: The second channel models the spatial correlation topology of environmental parameters through a graph neural network, further comprising: Map atmospheric environment parameters into dynamic graph nodes, and the node attributes include scalar parameters and vector parameters; Constructing connection edges between nodes based on the fluid dynamics relationship between meteorological elements, wherein the connection edges include connection weights between nodes; Perform multi-order graph convolution operations, where each order includes adjacency matrix regularization, degree matrix normalization, and feature nonlinear aggregation; Output is a graph embedding vector that characterizes the propagation characteristics of environmental disturbances.

5. The flight path stability optimization method according to claim 1, characterized in that: The policy network is trained using an actor-critic framework to achieve real-time closed-loop optimization of control actions, further including: When using a deep Q-network, the target network is combined with temporal difference learning to achieve iterative updates of the value function; When using proximal strategy optimization, the strategy update amplitude is constrained by importance sampling and gradient clipping; Establish a causal relationship model between control increment parameters and path stability; The contribution of attitude slope performance to path stability is verified through reward mechanism feedback.

6. The flight path stability optimization method according to claim 1, characterized in that: In response to the optimized control strategy, generating multi-parameter collaborative optimization instructions for flight control surfaces based on a real-time decision engine further includes: Deploy a lightweight inference engine to process fused feature streams in real time; Predict the combined impact of multi-control dimension adjustments on path stability; Generate a collaborative optimization instruction set including thrust compensation, rudder correction, roll compensation angle, and pitch compensation angle; A hierarchical warning and operation guidance mechanism is triggered based on the changing trend of the reward function.

7. A flight path stability optimization system, characterized in that: include: a data acquisition module configured to construct a dynamic fusion data set by aligning timestamps in response to aircraft dynamic parameters, pilot physiological state parameters, and three-dimensional space environment parameters collected synchronously from multiple sources; an architecture extraction module configured to extract dual-channel features of the dynamic fusion dataset based on a pre-built dual-channel feature extraction architecture in response to the dynamic fusion dataset, wherein the first channel analyzes the dynamic coupling relationship in the time dimension through a sparse attention mechanism, and the second channel models the spatial correlation topology of environmental parameters through a graph neural network; a strategy optimization module configured to fuse the dual-channel features to construct a deep reinforcement learning decision engine, establish a mapping strategy between environment state, control action, and path stability, drive control strategy optimization through a reward mechanism, and construct a multi-dimensional hybrid feature space as an environment state representation, including the temporal dynamic features output by the first channel and the spatial topological features output by the second channel; Define a four-dimensional continuous action space as the control action output, corresponding to the control increments of the power system, control surface, roll axis and pitch axis respectively; Define a multi-objective reward function, including path deviation penalty, speed fluctuation penalty, and stability gain reward; Adopting an actor-critic framework to train the policy network and achieve real-time closed-loop optimization of control actions; The decision suggestion module is configured to generate multi-parameter collaborative optimization instructions for the flight control surfaces based on the real-time decision engine in response to the optimized control strategy to improve the path stability.

8. 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; A computer program is stored in the memory, and when the computer program is executed by the processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 6.

9. 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 6.

Images