An aircraft pitch moment control method, system and apparatus

Abstract

The application provides an aircraft pitching moment control method, system and device, and relates to the technical field of aircraft pitching moment control. The method comprises the following steps: collecting pressure data and actual pitching moment of each detection point under gust disturbance; calculating an actual pitching moment coefficient curve and constructing an ideal pitching moment coefficient curve; extracting spatial features through a graph neural network based on the geometric coordinates and pressure data of each detection point; performing sequence modeling on a sliding time window according to a time sequence-based Transformer encoder; taking the difference between the actual pitching moment prediction value and the ideal pitching moment prediction value as a tracking error, outputting a control instruction increment through a PID controller, and then outputting a rudder deflection angle command through an actuator. The scheme can accurately capture and describe the nonlinear lift, drag and pitching moment coefficient changes under the aerodynamic disturbance environment, especially the aerodynamic dynamics under the conditions of large attack angle, strong nonlinearity and turbulence.

Description

Technical Field

[0001] This invention relates to the field of aircraft pitch moment control technology, and in particular to an aircraft pitch moment control method, system and device. Background Technology

[0002] As high-performance aircraft are increasingly used in complex aerodynamic environments, the impact of external disturbances such as gusts on their aerodynamic characteristics and attitude stability is becoming more and more prominent. Especially under transonic and high angle-of-attack flight conditions, transient pressure fluctuations caused by gusts can significantly affect the pitch moment of the aircraft, thereby increasing the attitude response deviation and reducing flight stability.

[0003] Traditional aerodynamic databases are mostly based on static wind tunnel tests or numerical simulations, which not only incur huge computational costs but also struggle to accurately capture unsteady aerodynamic characteristics under gust disturbances. While simplified aerodynamic models offer improved computational efficiency, they often lack accuracy under strongly nonlinear disturbances, making it difficult to meet both engineering applications and control law design requirements.

[0004] Therefore, establishing high-precision, real-time aerodynamic prediction and control methods has become an important direction in the research of integrated aerodynamics and flight control for aircraft. To this end, some targeted methods have been developed, but they still have varying degrees of problems: Simplified modeling methods based on traditional aerodynamic derivatives construct aerodynamic models through linearization or low-order approximations, which are suitable for flight control law design and rapid simulation, but are difficult to reflect the nonlinear effects under gusts. CFD-based numerical simulation methods utilize high-fidelity flow field calculations to obtain pressure and aerodynamic changes under gusts, which are suitable for mechanism analysis. However, the calculation time is still very long and the real-time performance is poor. While machine learning methods (such as LSTM, GRU, Transformer, etc.) can improve accuracy and adaptability to some extent by directly predicting aerodynamic coefficients from aerodynamic pressure sensor data, they usually ignore the spatial correlation of pressure distribution. Existing models cannot fully utilize the geometric configuration information of sensor distribution, resulting in poor prediction stability and model generalization. Moreover, they have "black box" characteristics, lacking physical constraints and interpretability. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, and device for controlling the pitch moment of an aircraft, so as to solve at least one of the above-mentioned technical problems existing in the prior art.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a method for controlling the pitch moment of an aircraft, comprising the following steps: Step 1: Deploy [measurements] on the surface of the aircraft to be tested. Each pressure sensor is used to obtain corresponding detection points; (based on wind tunnel or flight tests) pressure data and actual pitching moment are collected at each detection point under gust disturbances. ; Calculate the actual pitching moment coefficient curve Construct the ideal pitching moment coefficient curve .

[0007] In one feasible implementation, the pressure sensors are deployed at the leading edge, middle chord, trailing edge, and wingtip of the wing to improve observability and robustness.

[0008] In one feasible implementation, the specific data acquisition method in step 1 includes: Step 11: According to the sampling period The surface pressure at each detection point was obtained by synchronously collecting and calibrating the data. Specifically, the original electrical signal of surface pressure can be converted into a surface pressure value through calibration methods such as static pressure / zero point / temperature drift; the pressure coefficient can be calculated, and the specific expression includes: ; in, Indicates the incoming static pressure; Indicates dynamic pressure; Indicates air density; Indicates the incoming flow velocity; Indicates the first Surface pressure coefficient at each detection point; Step 12, for Denoising, normalization, and sliding window slicing are performed to obtain time-series input data for subsequent construction of ideal pitch moment and network training; Step 13: (Based on wind tunnel testing using a six-component force balance or force and moment measurement system) Synchronously collect the lift of the aircraft. ,resistance and pitching moment The lift coefficient curve was calculated. Drag coefficient curve and actual pitching moment coefficient curve The specific expressions include: ; in, Indicates the reference area. Indicates the reference chord length, used for dimensionless coordinates.

[0009] In one feasible implementation, the specific method for constructing the ideal pitching moment curve in step 1 includes: Step 14: Within the stable platform range before the disturbance, perform disturbance segment start-up detection.

[0010] In one feasible implementation, the specific detection method for the starting point of the disturbance segment includes: Step 141: Calculate the baseline values ​​and fluctuation ranges of lift, drag, and pitch moment. Specific expressions include: ; in, This indicates the statistical window of the stable platform before the disturbance. Inside The baseline value; express Inside The fluctuation range; express The starting time point; express The end time point; Step 142, when Exceeding a preset threshold (e.g.) times When the rate of change of deviation continues to increase, the current time point is determined to be the starting time point of the disturbance segment. .

[0011] Step 15: Within the stabilization plateau range after the disturbance, perform the end-point detection of the disturbance segment.

[0012] In one feasible implementation, the specific detection method for the end point of the disturbance segment includes: Step 151: Find the first time point when the deviation rate of change is zero and remains stable, and take it as the end point of the disturbance segment. ; Step 152: If it is still not found Then it will be continuous This occurred at a specific time point. Monotonically decreasing or The first time point when the moving average falls below a preset threshold is used as the time point when the threshold is first met. .

[0013] Step 16: Between the start and end points of the disturbance segment, the lift, drag, and pitch moment coefficient curves are smoothly reconstructed using an interpolation algorithm (e.g., cubic spline interpolation or fitting with smoothing constraints) to ensure the smoothness of the signal when entering and exiting the disturbance, while avoiding excessively sharp or violent fluctuations. Step 17: If any curve exhibits overshoot exceeding expectations, gradually reduce the slope at the endpoints (i.e., the boundaries of the disturbance segment) to allow the curve shape to converge naturally, resulting in a (smooth and physically reasonable) disturbance segment curve, which serves as the ideal pitching moment coefficient curve. The specific expression of the endpoint includes: and .

[0014] It should be noted that, and These are all dimensionless coefficients, not pressure values.

[0015] In one feasible implementation, step 18 may also be included. Add to the original data table to generate actual-ideal comparison curves (for each group of wind tunnel tests) for visual inspection and verification.

[0016] Step 2: Based on the geometric coordinates and pressure data of each detection point, spatial features are extracted using a graph neural network (GAT) to model the spatial coupling relationship between pressure distributions, thereby obtaining the first prediction model and capturing the spatial correlation of unsteady aerodynamic loads.

[0017] In one feasible implementation, the method for constructing the graph neural network in step 2 includes: Step 21: According to the length of the first sliding time window Construct samples, and at each time step (sample the time points), the th... The surface pressure coefficient and geometric coordinates of each detection point form the node feature vector. and its input tensor And there are The specific expressions include: ; in, , Indicates the first Geometric coordinates of each detection point; Step 22: Treat each pressure sensor as a node in the graph and construct the graph (using a fully connected graph or an edge connection method based on a distance threshold or nearest neighbor). This is to characterize the spatial coupling relationship between each detection point; among which, This represents the set of nodes consisting of the corresponding nodes of each pressure sensor; This represents the set of edges established between nodes according to preset connection rules; Step 23: Introduce radial basis function (RBF) similarity based on coordinate distance into the graph as a logarithmic bias for attention, so that similar detection points receive higher weights during attention calculation. The specific expression includes: ; in, Indicates the first The dimensionless coordinates of each node, i.e. ; Represents (settable or learnable) scale parameters; Indicates the first The node and the first Radial basis similarity between nodes; express Log-biased attention; Step 24: At each time step, perform a linear mapping on the node feature vectors and calculate the attention weights. The specific expressions include: ; ; ; in, Indicates the first The node and the first Attention score between nodes; Indicates the first The node is the first Normalized attention weights for each node; and Indicates trainable parameters; Represents (configurable or learnable) prior weight coefficients used to adjust the strength of the influence of coordinate prior bias on attention calculation; Indicates the first The node embedding vector obtained by neighborhood aggregation of each node; Indicates the first The set of neighboring nodes of a node; This represents vector concatenation; in this process of fusing priors, multi-head attention and residual connections can be used to enhance expressive power and training stability. Step 25: Embed all nodes Then, attention-weighted pooling is performed to obtain the global spatial feature vector at that time step. and its time step vector sequence The specific expressions include: ; in, Indicates the first Normalized attention pooling weights for each node (during global spatial feature aggregation).

[0018] In one feasible implementation, when some pressure sensors have missing readings or abnormal values, threshold detection and denoising can be performed first, and then repaired by methods such as linear interpolation, forward padding, or weighted estimation based on neighboring sensors. At the same time, a missing reading mask is introduced as an additional input to enhance the robustness of the model to the missing reading problem.

[0019] In one feasible implementation, when the actual geometric coordinates of the pressure sensor are temporarily unavailable, they can be replaced by predefined template coordinates (i.e., placeholder coordinates). To ensure the model can run, after obtaining the real geometric coordinates, the template coordinates are replaced, and retraining or incremental calibration is performed.

[0020] Step 3: Based on the first prediction model, and using a time-based Transformer encoder, perform sequence modeling on the sliding time window to obtain the second prediction model; the second prediction model is used to... and The actual pitch moment prediction and the ideal pitch moment prediction are obtained through reasoning, so as to learn the dynamic image of gust disturbance on lift, drag and pitch moment, and realize multi-task synchronous regression.

[0021] In one feasible implementation, step 3 specifically includes: Step 31: Define the output vector The specific expressions include: ; Step 32, (for each wind tunnel test sequence) according to the length of the second sliding time window (e.g., 128 sampling points) and step size (For example, 4 sampling points), construct training samples; define the input vector sequence of the training samples, the specific expression of which includes: ; in, Indicates the first The input vector sequence of training samples; express Feature dimensions; The end of the sliding time window (or the joint regression of multiple steps within the sliding time window) is used as the output label. The specific expression includes: ; in, Indicates the first The output labels of each training sample; Step 33: Overlay positional encoding onto the global spatial feature vector to introduce temporal positional information. The specific expression includes: ; in, Represents the feature embedding matrix; Indicates the first The position code corresponding to each sequence position can be a sine or cosine position code or a learnable position code. Indicates the sequence position index within the second sliding time window; Indicates the first The input embedding vector for each sequence position; Will The input to the multi-layer Transformer encoder (which is constructed using a multi-head self-attention network combined with a feedforward network as a single-layer structure, along with residual connections and layer normalization) includes the following specific expressions: ; in, The intermediate feature representation (after processing by the multi-head self-attention module and completing residual connections and layer normalization); This represents the encoded output (after processing by the feedforward network and completing residual connections and layer normalization); This indicates a multi-head self-attention module; Indicates a feedforward network; Representation layer normalization; Step 34: Add (learnable) convergence markers (CLS) before the sequence input to the multi-layer Transformer encoder, and take the convergence marker vector of the final layer as the first... Sequence representation of training samples ;Will The data is fed into the multilayer sensor regression head, which outputs predicted values ​​for the lift coefficient, drag coefficient, actual pitch moment coefficient, and ideal pitch moment coefficient. Specific expressions include: ; in, Indicates the first Predicted values ​​for each training sample; Represents a multilayer perceptron; Step 35: Use the weighted mean squared error as the training loss. The specific expressions include: ; in, , , and These represent the task weights corresponding to the lift coefficient, drag coefficient, actual pitch moment coefficient, and ideal pitch moment coefficient, respectively. , , and These represent the predicted values ​​of the lift coefficient, drag coefficient, actual pitch moment coefficient, and ideal pitch moment coefficient output by the multilayer sensor regression head (in step 34); , , and These represent the actual values ​​corresponding to the predicted values; Preferably, during the training phase of the second prediction model, the parameters are optimized using the AdamW optimizer, and a mixture of precision and gradient pruning methods can be used to improve stability. Step 36: In the inference stage of the second prediction model, the actual pitch moment prediction value, the ideal pitch moment prediction value, and the distribution of attention weights in the time dimension are output to characterize the model’s attention to different time segments, so as to help identify key stages such as disturbance triggering, disturbance peak and stabilization.

[0022] Step 4: The difference between the actual predicted pitch moment and the ideal predicted pitch moment is used as the tracking error. This error is then used by the PID controller to output the incremental control command. Then, the actuator outputs the control surface deflection angle command. .

[0023] In one feasible implementation, the PID controller includes a proportional term, an integral term, and a derivative term; the proportional term is used for fast response; the integral term is used to eliminate steady-state deviation; and the derivative term is used to suppress abrupt changes and overshoot. The PID controller also includes a conventional fixed-frequency update and zero-order hold strategy to suppress high-frequency jitter; The PID controller also includes conventional rudder deflection angle saturation and integral wind protection up strategies to avoid excessive commands; The PID controller also includes a conventional exponential smoothing strategy to balance response and stability; The PID controller also includes a conventional adaptive step size strategy to improve numerical robustness.

[0024] In one feasible implementation, the actuator includes at least one of the following three types: Linear actuators, specifically including the following expressions: ; in, Indicates actuator gain; Indicates the maximum deflection angle; This represents a saturation function, used to restrict values ​​between upper and lower bounds; Sine actuator, the specific expressions include: ; in, This represents the saturation angle scale parameter, used to characterize the steepness of the nonlinear limiting. Represents a symbolic function; Represents the sine function; Hyperbolic tangent actuator, the specific expression includes: ; in, Represents the hyperbolic tangent function; The sinusoidal actuator and the hyperbolic tangent actuator can more smoothly describe the nonlinear limiting of the actuator under large commands.

[0025] Secondly, based on the same inventive concept, this application also provides an aircraft pitch moment control system, including a data acquisition module, a data processing module and a result generation module; The data acquisition module is used to collect pressure data and actual pitching moment at each detection point under gust wind disturbance. ; The data processing module includes an ideal pitch moment unit, a first prediction model unit, a second prediction model unit, and a control execution unit. The ideal pitch moment unit is used to calculate the actual pitch moment coefficient curve. Construct the ideal pitching moment coefficient curve ; The first prediction model unit is used to extract spatial features based on the geometric coordinates and pressure data of each detection point, and to model the spatial coupling relationship between pressure distributions through a graph neural network to obtain the first prediction model. The second prediction model unit is used to perform sequence modeling on the sliding time window based on the first prediction model and according to the time-based Transformer encoder to obtain the second prediction model; the second prediction model is used to... and The predicted values ​​of the actual pitch moment and the ideal pitch moment are obtained through reasoning. The control execution unit is used to take the difference between the actual pitch moment prediction value and the ideal pitch moment prediction value as the tracking error, and output the control command increment through the PID controller. Then, the actuator outputs the control surface deflection angle command. ; The result generation module is used to... Outsourced.

[0026] Thirdly, based on the same inventive concept, this application also provides an aircraft pitch moment control device, including a processor, a memory, and a bus. The memory stores instructions and data that can be read by the processor. The processor is used to call the instructions and data in the memory to execute the aircraft pitch moment control method as described above. The bus connects the various functional components for transmitting information.

[0027] By adopting the above technical solution, the present invention has the following beneficial effects: This invention provides a method, system, and device for controlling the pitch moment of an aircraft, which can accurately capture and describe the changes in nonlinear lift, drag, and pitch moment coefficients under aerodynamic disturbance environments, especially the aerodynamic dynamics under large angles of attack, strong nonlinearity, and turbulent conditions. This scheme, by explicitly introducing sensor coordinate information into the input features and constructing an ideal reference pitch moment as the control target, can effectively learn and approximate complex nonlinear mapping relationships, overcoming the problem of insufficient prediction accuracy of traditional linearization methods in strongly nonlinear scenarios.

[0028] This approach introduces a coordinate-based graph attention mechanism into the network structure, constructing a prior correlation matrix using the geometric distance between nodes to ensure that the learning process conforms to the spatial distribution of sensors. Simultaneously, it incorporates a temporal Transformer to extract long-term dependent features, guaranteeing the physical consistency and stability of the prediction results. This combination of "physical constraints + deep learning" enhances both the model's interpretability and its adaptability to non-stationary perturbations.

[0029] This scheme generates the desired pitch moment as a control reference signal in advance by using an independent ideal pitch moment unit. Unlike traditional methods that rely on experiments or numerical simulations to obtain the target, this scheme can flexibly adjust the control target during the design phase, thereby ensuring that the training model has a clear optimization direction and improving the overall control effect and practicality.

[0030] This scheme does not rely on a single aerodynamic condition or a specific operating condition. Instead, it constructs a dataset by batch grouping sliding window sequences, which can simultaneously cover aerodynamic environments with different angles of attack, different disturbance intensities, and different turbulence characteristics, and maintain high prediction accuracy and control performance under multiple operating conditions and multiple disturbances.

[0031] This approach maintains a low prediction error while significantly outperforming traditional numerical simulation methods in inference speed. Combining distributed training and batch inference, this approach achieves millisecond-level prediction and control updates, making it suitable for real-time closed-loop control requirements in high-speed dynamic scenarios. Attached Figure Description

[0032] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0033] Figure 1 A flowchart of an aircraft pitch moment control method provided in an embodiment of the present invention; Figure 2A comparison diagram between the actual pitching moment and the ideal pitching moment during the upward phase of gust disturbance, provided for an embodiment of the present invention; Figure 3 A comparison diagram between the actual pitching moment and the ideal pitching moment during the downswing phase of a gust of wind, provided in an embodiment of the present invention; Figure 4 The comparison charts are of prediction results without attention background provided in the embodiments of the present invention; wherein, chart a is a comparison of lift, chart b is a comparison of drag, and chart c is a comparison of pitch moment. Figure 5 The comparison charts of prediction results with attention background provided in the embodiments of the present invention are shown; wherein, chart a is a comparison of lift, chart b is a comparison of drag, and chart c is a comparison of pitch moment. Figure 6 A schematic diagram of the distribution of 16 pressure sensors on a triangular wing surface provided in an embodiment of the present invention; Figure 7 The diagram shows the control effect of the linear actuator provided in the embodiment of the present invention; wherein, Figure a is a comparison diagram between the predicted actual pitch moment and the ideal pitch moment; Figure b is a comparison diagram between the actual pitch moment, the total pitch moment after control, and the reference ideal pitch moment; Figure c is a curve showing the change of the control surface deflection angle over time; Figure 8 The diagram shows the control effect of the sinusoidal actuator provided in the embodiment of the present invention; wherein, Figure a is a comparison diagram between the predicted actual pitch moment and the ideal pitch moment; Figure b is a comparison diagram between the actual pitch moment, the total pitch moment after control, and the reference ideal pitch moment; Figure c is a curve showing the change of the control surface deflection angle over time; Figure 9 The diagram shows the control effect of the hyperbolic tangent actuator provided in the embodiment of the present invention; wherein, Figure a is a comparison diagram between the predicted actual pitch moment and the ideal pitch moment; Figure b is a comparison diagram between the actual pitch moment, the total pitch moment after control, and the reference ideal pitch moment; Figure c is a curve showing the change of the control surface deflection angle over time; Figure 10 A diagram of an aircraft pitch moment control system provided for an embodiment of the present invention. Detailed Implementation

[0034] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0036] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0037] The present invention will be further explained below with reference to specific embodiments.

[0038] It should also be noted that the specific embodiments or implementation methods described below are a series of optimized settings listed by the present invention to further explain the specific content of the invention, and these settings can be combined or used in conjunction with each other.

[0039] Example 1: like Figure 1 As shown in the figure, the aircraft pitch moment control method provided in this embodiment includes the following steps: Step 1: Deploy test equipment on the surface of the aircraft to be tested (e.g., the critical aerodynamic surfaces of the wing / fuselage). Each pressure sensor is used to obtain corresponding detection points; based on wind tunnel tests, pressure data and actual pitching moment at each detection point are collected under gust wind disturbances. ; Calculate the actual pitching moment coefficient curve Construct the ideal pitching moment coefficient curve .

[0040] Furthermore, the pressure sensors are specifically deployed at the leading edge, middle chord, trailing edge, and wingtip of the wing to improve observability and robustness.

[0041] Furthermore, the specific data acquisition method in step 1 includes: Step 11: According to the sampling period The surface pressure at each detection point was obtained by synchronously collecting and calibrating the data. Specifically, the original electrical signal of surface pressure can be converted into a surface pressure value through calibration methods such as static pressure / zero point / temperature drift; the pressure coefficient can be calculated, and the specific expression includes: ; in, Indicates the incoming static pressure; Indicates dynamic pressure; Indicates air density; Indicates the incoming flow velocity; Indicates the first Surface pressure coefficient at each detection point; Step 12, for Denoising, normalization, and sliding window slicing are performed to obtain time-series input data for subsequent construction of ideal pitch moment and network training; Step 13: Based on wind tunnel testing, synchronously collect the lift of the aircraft (using a six-component force balance or a force and moment measurement system). ,resistance and pitching moment The lift coefficient curve was calculated. Drag coefficient curve and actual pitching moment coefficient curve The specific expressions include: ; in, Indicates the reference area. Indicates the reference chord length, used for dimensionless coordinates.

[0042] Furthermore, the specific method for constructing the ideal pitching moment curve in step 1 includes: Step 14: Within the stable platform range before the disturbance, perform disturbance segment start-up detection.

[0043] Furthermore, the specific detection method for the starting point of the disturbance segment includes: Step 141: Calculate the baseline values ​​and fluctuation ranges of lift, drag, and pitch moment. Specific expressions include: ; in, This indicates the statistical window of the stable platform before the disturbance. Inside The baseline value; express Inside The fluctuation range; express The starting time point; express The end time point; Step 142, when Exceeding a preset threshold (e.g.) times When the rate of change of deviation continues to increase, the current time point is determined to be the starting time point of the disturbance segment. .

[0044] Step 15: Within the stabilization plateau range after the disturbance, perform the end-point detection of the disturbance segment.

[0045] Furthermore, the specific detection method for the end point of the disturbance segment includes: Step 151: Find the first time point when the deviation rate of change is zero and remains stable, and take it as the end point of the disturbance segment. ; Step 152: If it is still not found Then it will be continuous This occurred at a specific time point. Monotonically decreasing or The first time point when the moving average falls below a preset threshold is used as the time point when the threshold is first met. .

[0046] Step 16: Between the start and end points of the disturbance segment, the lift, drag, and pitch moment coefficient curves are smoothly reconstructed using an interpolation algorithm (e.g., cubic spline interpolation or fitting with smoothing constraints) to ensure the smoothness of the signal when entering and exiting the disturbance, while avoiding excessively sharp or violent fluctuations. Step 17: If any curve exhibits overshoot exceeding expectations, gradually reduce the slope at the endpoints (i.e., the boundaries of the disturbance segment) to allow the curve shape to converge naturally, resulting in a (smooth and physically reasonable) disturbance segment curve, which serves as the ideal pitching moment coefficient curve. The specific expression of the endpoint includes: and ; Step 18, Add to the original data table to generate actual-ideal comparison curves for each group of wind tunnel tests for inspection and verification.

[0047] It should be noted that, and These are all dimensionless coefficients, not pressure values.

[0048] According to experimental tests, such as Figure 2-3 As shown, regardless of whether it is the upward or downward phase of a gust disturbance, this scheme can automatically identify and generate a continuous and smooth ideal pitch moment curve within the disturbance range, while maintaining physical rationality and stability, making it more suitable as a target for subsequent control; for example, Figure 2 middle, The sampling point number (520 and 871) is represented by m, which represents the deviation between the actual pitch moment and the predicted pitch moment at the corresponding sampling point (0.000637 and 0.00097). Figure 3 middle, The sampling point number is represented by 520 and 834, and m represents the deviation between the actual pitch moment and the predicted pitch moment at the corresponding sampling point (-0.000695 and -0.000128).

[0049] Step 2: Based on the geometric coordinates and pressure data of each detection point, spatial features are extracted using a graph neural network (GAT) to model the spatial coupling relationship between pressure distributions, thereby obtaining the first prediction model and capturing the spatial correlation of unsteady aerodynamic loads.

[0050] Furthermore, the method for constructing the graph neural network in step 2 includes: Step 21: According to the length of the first sliding time window Construct samples, at each time step, the first... The surface pressure coefficient and geometric coordinates of each detection point form the node feature vector. and its input tensor And there are The specific expressions include: ; in, , Indicates the first Geometric coordinates of each detection point; Step 22: Treat each pressure sensor as a node in the graph, and construct the graph using a fully connected graph or based on a distance threshold (or a nearest neighbor connection method). This is to characterize the spatial coupling relationship between each detection point; among which, This represents the set of nodes consisting of the corresponding nodes of each pressure sensor; This represents the set of edges established between nodes according to preset connection rules; Step 23: Introduce radial basis function (RBF) similarity based on coordinate distance into the graph as a logarithmic bias for attention, so that similar detection points receive higher weights during attention calculation. The specific expression includes: ; in, Indicates the first The dimensionless coordinates of each node, i.e. ; Represents (settable or learnable) scale parameters; Indicates the first The node and the first Radial basis similarity between nodes; express Log-biased attention; Step 24: At each time step, perform a linear mapping on the node feature vectors and calculate the attention weights. The specific expressions include: ; ; ; in, Indicates the first The node and the first Attention score between nodes; Indicates the first The node is the first Normalized attention weights for each node; and Indicates trainable parameters; Represents (configurable or learnable) prior weight coefficients used to adjust the strength of the influence of coordinate prior bias on attention calculation; Indicates the first The node embedding vector obtained by neighborhood aggregation of each node; Indicates the first The set of neighboring nodes of a node; This represents vector concatenation; in this process of fusing priors, multi-head attention and residual connections can be used to enhance expressive power and training stability. Step 25: Embed all nodes Then, attention-weighted pooling is performed to obtain the global spatial feature vector at that time step. and its time step vector sequence The specific expressions include: ; in, Indicates the first Normalized attention pooling weights for each node during the global spatial feature aggregation process.

[0051] Furthermore, when some pressure sensors have missing readings or abnormal values, threshold detection and denoising can be performed first, and then repaired by methods such as linear interpolation, forward padding, or weighted estimation based on neighboring sensors. At the same time, a missing reading mask is introduced as an additional input to enhance the model's robustness to missing reading problems.

[0052] Furthermore, when the actual geometric coordinates of the pressure sensor are temporarily unavailable, predefined template coordinates (i.e., placeholder coordinates) can be used as a substitute. To ensure the model can run, after obtaining the real geometric coordinates, the template coordinates are replaced, and retraining or incremental calibration is performed.

[0053] Step 3: Based on the first prediction model, and using a time-based Transformer encoder, perform sequence modeling on the sliding time window to obtain the second prediction model; the second prediction model is used to... and The actual pitch moment prediction and the ideal pitch moment prediction are obtained through reasoning, so as to learn the dynamic image of gust disturbance on lift, drag and pitch moment, and realize multi-task synchronous regression.

[0054] Furthermore, step 3 specifically includes: Step 31: Define the output vector The specific expressions include: ; Step 32: For each wind tunnel test sequence, according to the length of the second sliding time window... (e.g., 128 sampling points) and step size (For example, 4 sampling points), construct training samples; define the input vector sequence of the training samples, the specific expression of which includes: ; in, Indicates the first The input vector sequence of training samples; express Feature dimensions; The end of the sliding time window (or the joint regression of multiple steps within the sliding time window) is used as the output label. The specific expression includes: ; in, Indicates the first The output labels of each training sample; Step 33: Overlay positional encoding onto the global spatial feature vector to introduce temporal positional information. The specific expression includes: ; in, Represents the feature embedding matrix; Indicates the first The position code corresponding to each sequence position can be a sine or cosine position code or a learnable position code. Indicates the sequence position index within the second sliding time window; Indicates the first The input embedding vector for each sequence position; Will The input is fed into a multi-layer Transformer encoder, which uses a multi-head self-attention network combined with a feedforward network as a single-layer structure, and is further configured with residual connections and layer normalization. The specific expression includes: ; in, The intermediate feature representation (after processing by the multi-head self-attention module and completing residual connections and layer normalization); This represents the encoded output (after processing by the feedforward network and completing residual connections and layer normalization); This indicates a multi-head self-attention module; Indicates a feedforward network; Representation layer normalization; Step 34: Add learnable convergence markers (CLS) before the sequence input to the multi-layer Transformer encoder, and take the convergence marker vector of the final layer as the first... Sequence representation of training samples ;Will The data is fed into the multilayer sensor regression head, which outputs predicted values ​​for the lift coefficient, drag coefficient, actual pitch moment coefficient, and ideal pitch moment coefficient. Specific expressions include: ; in, Indicates the first Predicted values ​​for each training sample; Represents a multilayer perceptron; Step 35: Use the weighted mean squared error as the training loss. The specific expressions include: ; in, , , and These represent the task weights corresponding to the lift coefficient, drag coefficient, actual pitch moment coefficient, and ideal pitch moment coefficient, respectively. , , and These represent the predicted values ​​of the lift coefficient, drag coefficient, actual pitch moment coefficient, and ideal pitch moment coefficient output by the multilayer sensor regression head (in step 34); , , and These represent the actual values ​​corresponding to the predicted values; During the training phase of the second prediction model, the parameters are optimized using the AdamW optimizer, and a mixture of precision and gradient clipping methods can be used to improve stability. Step 36: In the inference stage of the second prediction model, the actual pitch moment prediction value, the ideal pitch moment prediction value, and the distribution of attention weights in the time dimension are output to characterize the model’s attention to different time segments, so as to help identify key stages such as disturbance triggering, disturbance peak and stabilization.

[0055] Experimental tests showed that, in the absence of attention, such as Figure 4 As shown in Figure a, the predicted lift curve of this scheme is basically consistent with the actual lift curve; as... Figure 4 As shown in Figure b, the predicted resistance curve of this scheme is basically consistent with the actual resistance curve; as... Figure 4 As shown in Figure c, the predicted actual pitch moment curve of this scheme is basically consistent with the actual pitch moment curve, and the predicted ideal pitch moment curve is basically consistent with the ideal pitch moment curve. Similarly, in a context with attention, such as Figure 5 As shown in Figure a, the predicted lift curve of this scheme is basically consistent with the actual lift curve; as... Figure 5 As shown in Figure b, the predicted resistance curve of this scheme is basically consistent with the actual resistance curve; as... Figure 5 As shown in Figure c, the predicted actual pitch moment curve of this scheme is basically consistent with the actual pitch moment curve, and the predicted ideal pitch moment curve is basically consistent with the actual ideal pitch moment curve. In particular, the attention peak and the gust disturbance upsurge-peak-recovery stages are highly coincident, proving that the graph neural network based on coordinate prior combined with the time-series based Transformer encoder in this scheme can spontaneously focus on key dynamic ranges, thereby significantly improving the prediction stability and interpretability for sudden and decaying processes.

[0056] Step 4: The difference between the actual predicted pitch moment and the ideal predicted pitch moment is used as the tracking error. This error is then used by the PID controller to output the incremental control command. Then, the actuator outputs the control surface deflection angle command. .

[0057] Furthermore, the PID controller includes a proportional term, an integral term, and a derivative term; the proportional term is used for fast response; the integral term is used to eliminate steady-state deviation; and the derivative term is used to suppress abrupt changes and overshoot. The PID controller also includes a conventional fixed-frequency update and zero-order hold strategy to suppress high-frequency jitter; The PID controller also includes conventional rudder deflection angle saturation and integral wind protection up strategies to avoid excessive commands; The PID controller also includes a conventional exponential smoothing strategy to balance response and stability; The PID controller also includes a conventional adaptive step size strategy to improve numerical robustness.

[0058] Furthermore, the actuator includes at least the following three types: Linear actuators, specifically including the following expressions: ; in, Indicates actuator gain; Indicates the maximum deflection angle; This represents a saturation function, used to restrict values ​​between upper and lower bounds; Sine actuator, the specific expressions include: ; in, This represents the saturation angle scale parameter, used to characterize the steepness of the nonlinear limiting. Represents a symbolic function; Represents the sine function; Hyperbolic tangent actuator, the specific expression includes: ; in, Represents the hyperbolic tangent function; The sinusoidal actuator and the hyperbolic tangent actuator can more smoothly describe the nonlinear limiting of the actuator under large commands.

[0059] For example, to verify the effectiveness of this solution, 16 pressure sensors were deployed on the surface of the aircraft's delta wing model. Their specific locations and numbering are as follows: Figure 6 As shown, a wind tunnel experiment was then conducted to collect surface pressure data that varied over time. After normalization and geometric coordinate encoding, the data was input into the trained second prediction model for comparative experiments.

[0060] Experimental control objectives: structure And compare it with the predicted value of the ideal pitching moment; Experimental comparison scheme: Traditional linear PID control and traditional nonlinear (sine / hyperbolic tangent) PID control schemes; Experimental evaluation indicators: Mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), and real-time prediction performance; Specific experimental results: like Figure 7-9The results shown illustrate the experimental outcomes of control execution using linear actuators, sinusoidal actuators, and hyperbolic tangent actuators under the 28th set of timing input data. for Figure 7 : PID parameters include: ;in, Indicates proportional gain; Indicates integral gain; Represents differential gain; Indicates the linear actuator gain, used to make... Linear mapping is ; Figure 7 Figure a shows the comparison between the predicted actual pitch moment and the ideal pitch moment under linear actuator; Figure b shows the comparison between the actual pitch moment, the total pitch moment after control, and the reference ideal pitch moment under linear actuator; Figure c shows the curve of the control surface deflection angle changing with time under linear actuator. for Figure 8 : PID parameters include: ; Figure 8 Figure a shows the comparison between the predicted actual pitch moment and the ideal pitch moment under the sinusoidal actuator; Figure b shows the comparison between the actual pitch moment, the total pitch moment after control, and the reference ideal pitch moment under the sinusoidal actuator; Figure c shows the curve of the control surface deflection angle changing with time under the sinusoidal actuator. for Figure 9 : PID parameters include: ; Figure 9 Figure a shows the comparison between the predicted actual pitch moment and the ideal pitch moment under the hyperbolic tangent actuator; Figure b shows the comparison between the actual pitch moment, the total pitch moment after control, and the reference ideal pitch moment under the hyperbolic tangent actuator; Figure c shows the curve of the control surface deflection angle changing with time under the hyperbolic tangent actuator. As shown in the figure, the predicted pitch moment and the ideal value of the configuration are well consistent under all three actuators. For the linear actuator, the total pitch moment coefficient after control can be expressed as: ; in, This represents the actual pitch moment coefficient before control. Indicates the gain of the linear actuator; Indicates the control surface deflection angle command; This represents the total pitch moment coefficient after control. The controlled total pitch moment can effectively track the reference curve. At the same time, the control surface deflection angle changes smoothly, indicating that all three actuators can effectively control the pitch moment under gust disturbances, thus ensuring the stability and feasibility of the flight control system. This scheme improves the adaptability to nonlinear actuators and different gust disturbance conditions while ensuring prediction accuracy, and is suitable for real-time aerodynamic characteristic prediction and active control of aircraft.

[0061] Example 2: like Figure 10 As shown, this embodiment provides an aircraft pitch moment control system, including a data acquisition module, a data processing module, and a result generation module; The data acquisition module is used to collect pressure data and actual pitching moment at each detection point under gust wind disturbance. ; The data processing module includes an ideal pitch moment unit, a first prediction model unit, a second prediction model unit, and a control execution unit. The ideal pitch moment unit is used to calculate the actual pitch moment coefficient curve. Construct the ideal pitching moment coefficient curve ; The first prediction model unit is used to extract spatial features based on the geometric coordinates and pressure data of each detection point, and to model the spatial coupling relationship between pressure distributions through a graph neural network to obtain the first prediction model. The second prediction model unit is used to perform sequence modeling on the sliding time window based on the first prediction model and according to the time-based Transformer encoder to obtain the second prediction model; the second prediction model is used to... and The predicted values ​​of the actual pitch moment and the ideal pitch moment are obtained through reasoning. The control execution unit is used to take the difference between the actual pitch moment prediction value and the ideal pitch moment prediction value as the tracking error, and output the control command increment through the PID controller. Then, the actuator outputs the control surface deflection angle command. ; The result generation module is used to... Outsourced.

[0062] Example 3: This embodiment provides an aircraft pitch moment control device, including a processor, a memory, and a bus. The memory stores instructions and data that can be read by the processor. The processor is used to call the instructions and data in the memory to execute the aircraft pitch moment control method as described above. The bus connects the various functional components for transmitting information.

[0063] In another embodiment, this solution can also be implemented using an integrated device, which may include corresponding modules that perform one or more steps in the various embodiments described above. A module may be one or more hardware modules specifically configured to perform the corresponding step, or implemented by a processor configured to perform the corresponding step, or stored in a computer-readable medium for implementation by a processor, or implemented through some combination thereof.

[0064] The processor executes the various methods and processes described above. For example, the method implementations in this scheme can be implemented as software programs tangibly contained in a machine-readable medium, such as memory. In some implementations, part or all of the software program can be loaded and / or installed via memory and / or a communication interface. When the software program is loaded into memory and executed by the processor, one or more steps of the methods described above can be performed. Alternatively, in other implementations, the processor can be configured to execute one of the methods described above by any other suitable means (e.g., by means of firmware).

[0065] This device can be implemented using a bus architecture. A bus architecture can include any number of interconnect buses and bridges, depending on the specific application of the hardware and overall design constraints. The bus connects various circuits, including one or more processors, memory, and / or hardware modules. The bus can also connect various other circuits such as peripherals, voltage regulators, power management circuitry, external antennas, etc.

[0066] Buses can be Industry Standard Architecture (ISA) buses, Peripheral Component Interconnect (PCI) buses, or Extended Industry Standard Component (EISA) buses, etc. Buses can be divided into address buses, data buses, control buses, etc.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions 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 method for controlling the pitch moment of an aircraft, characterized in that, include: Step 1: Deploy [measurements] on the surface of the aircraft to be tested. Each pressure sensor is used to obtain corresponding detection points; pressure data and actual pitching moment at each detection point are collected under gust wind disturbance. ; Calculate the actual pitching moment coefficient curve Construct the ideal pitching moment coefficient curve ; Step 2: Based on the geometric coordinates and pressure data of each detection point, spatial features are extracted using a graph neural network to model the spatial coupling relationship between pressure distributions, thus obtaining the first prediction model. Step 3: Based on the first prediction model, and using a time-based Transformer encoder, perform sequence modeling on the sliding time window to obtain the second prediction model; the second prediction model is used to... and The actual pitch moment prediction value and the ideal pitch moment prediction value are obtained by reasoning. Step 4: The difference between the actual predicted pitch moment and the ideal predicted pitch moment is used as the tracking error. This error is then used by the PID controller to output the incremental control command. Then, the actuator outputs the control surface deflection angle command. .

2. The control method according to claim 1, characterized in that, The pressure sensors are specifically installed at the leading edge, middle chord, trailing edge, and wingtip of the wing.

3. The control method according to claim 1, characterized in that, Data was collected through wind tunnel experiments.

4. The control method according to claim 1, characterized in that, The specific data acquisition method in step 1 includes: Step 11: According to the sampling period The surface pressure at each detection point was obtained by synchronously collecting and calibrating the data. The pressure coefficient is calculated using the following specific expressions: ； in, Indicates the incoming static pressure; Indicates dynamic pressure; Indicates air density; Indicates the incoming flow velocity; Indicates the first Surface pressure coefficient at each detection point; Step 12, for Denoising, normalization, and sliding window slicing are performed to obtain the time-series input data; Step 13: Synchronously collect the lift of the aircraft ,resistance and pitching moment The lift coefficient curve was calculated. Drag coefficient curve and actual pitching moment coefficient curve The specific expressions include: ； in, Indicates the reference area. Indicates the reference chord length.

5. The control method according to claim 1, characterized in that, The specific method for constructing the ideal pitching moment curve in step 1 includes: Step 14: Within the stable platform range before the disturbance, perform disturbance segment start-up detection; Step 15: Within the stabilization plateau range after the disturbance, perform end-point detection of the disturbance segment; Step 16: Between the start and end points of the disturbance segment, the lift, drag, and pitch moment coefficient curves are smoothly reconstructed using an interpolation algorithm. Step 17: If any curve shows an overshoot exceeding the expected value, gradually reduce the slope at the endpoints to obtain the disturbance segment curve, which serves as the ideal pitching moment coefficient curve. .

6. The control method according to claim 1, characterized in that, The method for constructing the graph neural network in step 2 includes: Step 21: According to the length of the first sliding time window Construct samples, at each time step, the first... The surface pressure coefficient and geometric coordinates of each detection point form the node feature vector. and its input tensor And there are The specific expressions include: ； in, , Indicates the first Geometric coordinates of each detection point; Step 22: Construct a graph by treating each pressure sensor as a node. ;in, This represents the set of nodes consisting of the corresponding nodes of each pressure sensor; This represents the set of edges established between nodes according to preset connection rules; Step 23: Introduce radial basis similarity based on coordinate distance into the graph as an attention log bias. The specific expression includes: ； in, Indicates the first The dimensionless coordinates of each node, i.e. ; Indicates the scale parameter; Indicates the first The node and the first Radial basis similarity between nodes; express Log-biased attention; Step 24: At each time step, perform a linear mapping on the node feature vectors and calculate the attention weights. The specific expressions include: ； ； ； in, Indicates the first The node and the first Attention score between nodes; Indicates the first The node is the first Normalized attention weights for each node; and Indicates trainable parameters; Indicates the prior weight coefficient; Indicates the first The node embedding vector obtained by neighborhood aggregation of each node; Indicates the first The set of neighboring nodes of a node; This indicates vector concatenation; Step 25: Embed all nodes Then, attention-weighted pooling is performed to obtain the global spatial feature vector at that time step. and its time step vector sequence The specific expressions include: ； in, Indicates the first Normalized attention pooling weights for each node.

7. The control method according to claim 1, characterized in that, Step 3 specifically includes: Step 31: Define the output vector The specific expressions include: ； Step 32: According to the length of the second sliding time window and step length Construct training samples; define the input vector sequence of the training samples, the specific expression of which includes: ； in, Indicates the first The input vector sequence of training samples; express Feature dimensions; The end of the sliding time window is used as the output label. The specific expression includes: ； in, Indicates the first The output labels of each training sample; Step 33: Superimpose positional encoding onto the global spatial feature vector. The specific expression includes: ； in, Represents the feature embedding matrix; Indicates the first The position code corresponding to each sequence position; Indicates the sequence position index within the second sliding time window; Indicates the first The input embedding vector for each sequence position; Will The input to the multi-layer Transformer encoder includes the following specific expressions: ； in, Indicates intermediate feature representation; Indicates the encoded output; This indicates a multi-head self-attention module; Indicates a feedforward network; Representation layer normalization; Step 34: Add a convergence marker before the sequence input to the multi-layer Transformer encoder, and take the convergence marker vector of the final layer as the first convergence marker. Sequence representation of training samples ;Will The data is fed into the multilayer sensor regression head, which outputs predicted values ​​for the lift coefficient, drag coefficient, actual pitch moment coefficient, and ideal pitch moment coefficient. Specific expressions include: ； in, Indicates the first Predicted values ​​for each training sample; This represents a multilayer perceptron; Step 35: Use the weighted mean squared error as the training loss. The specific expressions include: ； in, , , and These represent the task weights corresponding to the lift coefficient, drag coefficient, actual pitch moment coefficient, and ideal pitch moment coefficient, respectively. , , and These represent the predicted values ​​corresponding to the lift coefficient, drag coefficient, actual pitch moment coefficient, and ideal pitch moment coefficient output by the multilayer sensor regression head, respectively. , , and These represent the actual values ​​corresponding to the predicted values; Step 36: In the inference phase of the second prediction model, the outputs are the actual pitch moment prediction value, the ideal pitch moment prediction value, and the distribution of attention weights in the time dimension.

8. The control method according to claim 1, characterized in that, The actuator includes at least one of the following three types: Linear actuators, specifically including the following expressions: ； in, Indicates actuator gain; Indicates the maximum deflection angle; This represents a saturation function, used to restrict values ​​between upper and lower bounds; Sine actuator, the specific expressions include: ； in, Indicates the saturation angle scale parameter; Represents a symbolic function; Represents the sine function; Hyperbolic tangent actuator, the specific expression includes: ； in, This represents the hyperbolic tangent function.

9. A pitch moment control system for an aircraft, characterized in that, It includes a data acquisition module, a data processing module, and a result generation module; The data acquisition module is used to collect pressure data and actual pitching moment at each detection point under gust wind disturbance. ; The data processing module includes an ideal pitch moment unit, a first prediction model unit, a second prediction model unit, and a control execution unit. The ideal pitch moment unit is used to calculate the actual pitch moment coefficient curve. Construct the ideal pitching moment coefficient curve ; The first prediction model unit is used to extract spatial features based on the geometric coordinates and pressure data of each detection point, and to model the spatial coupling relationship between pressure distributions through a graph neural network to obtain the first prediction model. The second prediction model unit is used to perform sequence modeling on the sliding time window based on the first prediction model and according to the time-based Transformer encoder to obtain the second prediction model; the second prediction model is used to... and The predicted values ​​of the actual pitch moment and the ideal pitch moment are obtained through reasoning. The control execution unit is used to take the difference between the actual pitch moment prediction value and the ideal pitch moment prediction value as the tracking error, and output the control command increment through the PID controller. Then, the actuator outputs the control surface deflection angle command. ; The result generation module is used to... Outsourced.

10. A pitch moment control device for an aircraft, characterized in that, It includes a processor, a memory, and a bus. The memory stores instructions and data read by the processor. The processor is used to call the instructions and data in the memory to execute the control method as described in any one of claims 1-8. The bus connects the functional components for transmitting information.

Images