Swing nozzle decoupling control method and system of coupled neural network

Abstract

The invention discloses a swing nozzle decoupling control method and system of a coupled neural network. According to the scheme, attitude data of a spray pipe and actual expansion and contraction amount data of an actuator are collected through an optimized sampling strategy, the data are processed through a decoupling algorithm, and a spray pipe attitude-decoupling error data set is constructed; on the basis of the data set, a neural network model is trained to learn and predict decoupling errors of a decoupling algorithm under different nozzle postures at high precision; and in the control link, the theoretical output of the decoupling algorithm and error prediction of the neural network are coupled in real time, a high-precision actuator expansion and contraction amount control instruction is generated and sent to the actuator for execution, and precise control over the attitude of the spray pipe is achieved. According to the method, the attitude control precision of the swing nozzle is remarkably improved, accurate adjustment of the target swing angle is achieved through a single control instruction, the millisecond-level instantaneous response capability is achieved, and the dynamic performance of the swing nozzle is greatly improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of space engine thrust vector control, in particular to a swing nozzle decoupling control method and system coupled with a neural network. BACKGROUND

[0002] Thrust vector control technology is a key component in space engine systems, which adjusts the swing angle of the nozzle to achieve attitude adjustment and trajectory control of the aircraft, especially in high-speed flight and complex maneuvering tasks. The control accuracy of the swing nozzle as the actuator of thrust vector directly affects the stability and performance of the entire system.

[0003] Traditional control methods rely on decoupling algorithms to calculate the extension and retraction amount of the mechanical actuator under the preset attitude to drive the nozzle to reach the target attitude. However, due to the ideal simplification assumption of the decoupling algorithm, the machining and assembly errors of the mechanical structure, environmental disturbances and other factors, the decoupling algorithm often has inherent errors in actual application, leading to deviations in nozzle attitude control and affecting the accurate output of thrust vector. To avoid this deviation, the commonly used closed-loop control strategies such as PID control and sliding mode control cannot use a single control instruction to complete the accurate adjustment of the nozzle attitude due to the nature of iterative adjustment, making it difficult to meet the stringent requirements of high dynamic scenarios for millisecond-level instantaneous response. The existing compensation strategies based on physical models are trapped in the dilemma of strong model dependency, secondary deviation caused by simplification assumptions, and lack of adaptive ability to time-varying characteristics of the system. Therefore, it is urgent to develop a nozzle attitude control method with high control accuracy and fast response speed. SUMMARY

[0004] The application aims to provide a swing nozzle decoupling control method and system coupled with a neural network, which aims to overcome the control errors of the decoupling algorithm for the swing nozzle and improve the control accuracy and response speed.

[0005] To achieve the purpose of the application, on the one hand, the application provides a swing nozzle decoupling control method coupled with a neural network, which includes the following steps:

[0006] Step 1, collect nozzle attitude data and actual actuator extension and retraction data, process the data using a decoupling algorithm, and construct a nozzle attitude-decoupling error data set; the data set includes attitude data before and after the nozzle swings, and actuator extension and retraction error data calculated by the decoupling algorithm when the nozzle attitude after swinging is input;

[0007] Step 2, select the attitude data before and after the nozzle swings and the actuator extension and retraction error data of the decoupling algorithm as the input and output features of the decoupling error neural network, respectively, and normalize the feature data;

[0008] Step 3: Construct the decoupling error neural network model structure, determine the number of network layers and neurons, select the activation function, loss function and optimization algorithm type, and obtain the initial decoupling error neural network model;

[0009] Step 4: Divide the dataset into training set, validation set and test set, call the optimization algorithm to automatically optimize the hyperparameters of the neural network model, verify the effectiveness of the model, save the model and related parameters, and obtain the trained decoupling error neural network model.

[0010] Step 5: Call the saved decoupling error neural network model, input the nozzle target attitude parameters into the neural network and decoupling algorithm, perform forward propagation calculation of the neural network and solution of the decoupling algorithm, couple the error prediction of the neural network with the theoretical output of the decoupling algorithm in real time, generate actuator extension and retraction commands, and realize nozzle attitude control.

[0011] On the other hand, the present invention also provides a system for a decoupling control method for a swaying nozzle coupled with a neural network, comprising the following modules:

[0012] The data acquisition module is used to collect nozzle attitude data and actuator actual extension and retraction data, and to process the data using a decoupling algorithm to construct a nozzle attitude-decoupling error dataset. The dataset includes attitude data before and after nozzle swing, and actuator extension and retraction error data calculated by the decoupling algorithm when the attitude after nozzle swing is taken as input.

[0013] The data processing module is used to select the attitude data before and after nozzle swing and the actuator extension error data of the decoupling algorithm as the input and output features of the decoupling error neural network, and to normalize the feature data.

[0014] The decoupling error neural network model module is used to construct the decoupling error neural network model structure, determine the number of network layers and neurons, select the activation function, loss function and optimization algorithm type, and obtain the initial decoupling error neural network model.

[0015] The training module is used to divide the dataset into training set, validation set and test set, call optimization algorithm to automatically optimize the hyperparameters of the neural network model, verify the effectiveness of the model, save the model and related parameters, and obtain the trained decoupling error neural network model.

[0016] The extension / retraction command module is used to call the saved decoupling error neural network model, input the nozzle's target attitude parameters into the neural network and decoupling algorithm, perform forward propagation calculations of the neural network and decoupling algorithm calculations, couple the error prediction of the neural network with the theoretical output of the decoupling algorithm in real time, generate actuator extension / retraction commands, and realize nozzle attitude control.

[0017] The significant advancement of this invention compared to existing technologies lies in:

[0018] (1) This invention uses neural networks to accurately learn the error patterns of the decoupling algorithm under different nozzle attitudes, and couples its error prediction with the theoretical output of the decoupling algorithm in real time. This overcomes the inherent bias caused by the idealized assumptions of traditional models and can effectively fit complex nonlinear errors, thus significantly improving the attitude control accuracy of the oscillating nozzle from the source.

[0019] (2) This invention abandons the iterative convergence process of traditional closed-loop control and adopts a feedforward control strategy that uses decoupling algorithm and neural network parallel computing and result coupling. It can achieve precise adjustment of target swing angle with a single control command and has the instantaneous response capability at the millisecond level. It fundamentally solves the contradiction between speed and accuracy in high dynamic scenarios.

[0020] To more clearly illustrate the functional characteristics and structural parameters of the present invention, further explanation is provided below in conjunction with the accompanying drawings and specific embodiments. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0022] Figure 1 This is a flowchart of the steps of the present invention;

[0023] Figure 2 This is a flowchart of the present invention for constructing the nozzle attitude-decoupling error dataset;

[0024] Figure 3 This is a schematic diagram of the input and output features of the decoupling error neural network of the present invention;

[0025] Figure 4 This is a partial structural schematic diagram of the decoupling error neural network of the present invention;

[0026] Figure 5 This is a schematic diagram of the local coordinate system of the nozzle established by the decoupling algorithm of this invention;

[0027] Figure 6 This is a comparison diagram of the pitch angle command and measurement of the present invention;

[0028] Figure 7 This is a comparison diagram of the yaw angle command and measurement of the present invention;

[0029] Figure 8 This is a comparison diagram of the yaw angle and pitch angle errors of the present invention. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0031] This invention provides a decoupling control method for a coupled neural network-based oscillating nozzle, combining... Figure 1 This includes the following steps:

[0032] Step 1: Collect nozzle attitude data and actuator actual extension / retraction data, process the data using a decoupling algorithm, and construct a nozzle attitude-decoupling error dataset; the dataset includes attitude data before and after nozzle swing, and actuator extension / retraction error data calculated by the decoupling algorithm when the nozzle swing attitude is used as input.

[0033] Step 2: Select the attitude data before and after nozzle swing and the actuator extension error data of the decoupling algorithm as the input and output features of the decoupling error neural network, and normalize the feature data.

[0034] Step 3: Construct the decoupling error neural network model structure, determine the number of network layers and neurons, select the activation function, loss function and optimization algorithm type, and obtain the initial decoupling error neural network model;

[0035] Step 4: Divide the dataset into training set, validation set and test set, call the optimization algorithm to automatically optimize the hyperparameters of the neural network model, verify the effectiveness of the model, save the model and related parameters, and obtain the trained decoupling error neural network model.

[0036] Step 5: Call the saved decoupling error neural network model, input the nozzle target attitude parameters into the neural network and decoupling algorithm, perform forward propagation calculation of the neural network and solution of the decoupling algorithm, couple the error prediction of the neural network with the theoretical output of the decoupling algorithm in real time, generate high-precision actuator extension and retraction commands, and realize precise control of nozzle attitude.

[0037] Combination Figure 2 The construction of the nozzle attitude-decoupling error dataset in step 1 specifically includes the following steps:

[0038] Step 1-1: Generate a dataset of extension and retraction instructions based on the extension and retraction ranges of the two actuators, randomly shuffle the order of the instructions, and convert them into actuation commands.

[0039] Steps 1-2: Obtain the actual extension / retraction amount encoded inside the actuator according to the actuation command; obtain the nozzle attitude, including the nozzle pitch angle and yaw angle, through the attitude sensor;

[0040] Steps 1-3: Input the obtained nozzle attitude into the decoupling algorithm to calculate the theoretical actuator extension and retraction amount under the corresponding attitude;

[0041] Steps 1-4: Calculate the error of the decoupling algorithm based on the actual actuator extension / retraction amount obtained in Step 1-2 and the theoretical actuator extension / retraction amount obtained in Step 1-3;

[0042] Steps 1-5: Record the nozzle attitude acquired this time, the nozzle attitude acquired in the previous acquisition, and the scaling error of the decoupling algorithm. If it is the first acquisition, only record the nozzle attitude measured this time.

[0043] Steps 1-6: Repeat steps 1-2 to 1-5 until the collected data reaches the preset value. Construct a nozzle attitude-decoupling error dataset based on all collected data.

[0044] The scaling instruction dataset in step 1-1 is generated using a mixed uniform and random generation strategy, specifically as follows:

[0045] Select independently at equal intervals within the extension and retraction range of the two actuators. Generate points, and combine the generation points of one actuator with the generation points of another actuator in pairs to obtain... The actuator extension / retraction command; the random generation strategy is to randomly select within the extension / retraction range of the two actuators respectively. A number of generation points are generated by sequentially combining the generation points of one actuator with those of another. The actuator extension / retraction command; the uniform and random hybrid generation strategy is to use uniform generation and random generation to finally generate... The extension / retraction command of the actuator group; where... , , satisfy: .

[0046] The nozzle attitude input decoupling algorithm in steps 1-3 specifically includes the following steps:

[0047] Step 1-3-1: Establish a local coordinate system for the nozzle with the nozzle swing center as the origin, and determine the initial coordinates and initial distances of the connection points between the two actuators and the upper and lower nozzles.

[0048] Step 1-3-2: Determine the transformation matrix of the coordinates under the rotational motion of the nozzle, and calculate the coordinates of the follower point at the connection point between the two actuators and the nozzle after the nozzle swings based on the coordinate transformation matrix;

[0049] Step 1-3-3: Calculate the distance between the two actuators and the upper and lower connection points of the nozzle after the nozzle swings, and the extension and retraction of the two actuators in the corresponding state, based on the transformed coordinates.

[0050] The error calculation formula for the decoupling algorithm in steps 1-4 is as follows:

[0051] ;

[0052] in, This refers to the actual actuator extension or retraction in the direction of nozzle pitch or yaw angle. for The theoretical actuator extension / retraction amount in the corresponding direction. for The theoretical actuator extension / retraction error calculated by the decoupling algorithm in the corresponding direction.

[0053] The input and output features of the decoupling error neural network in step 2 are as follows:

[0054] Combination Figure 3 The input features are:

[0055] ;

[0056] in, This is the input vector of the neural network. The pitch angle characteristics before nozzle oscillation are measured by the attitude sensor. The yaw angle characteristics before nozzle oscillation are measured by the attitude sensor. The pitch angle characteristics after nozzle oscillation are measured by the attitude sensor. The yaw angle characteristics after nozzle wobbling, measured by an attitude sensor; wherein the elements of each characteristic satisfy:

[0057] ;

[0058] in, , The first The pitch and yaw angles before the nozzle oscillation were collected. , The first -1 acquisition of pitch and yaw angles after nozzle oscillation, among which , This represents the total number of data collections.

[0059] The output feature is:

[0060] ;

[0061] in, This is the output vector of the neural network. To determine the error characteristics of the theoretical actuator extension / retraction in the nozzle pitch direction calculated by the decoupling algorithm. The error characteristics of the actuator extension / retraction amount in the nozzle yaw angle direction calculated by the decoupling algorithm.

[0062] Preferably, min-max normalization is used to normalize the input and output features of the decoupling error neural network, and the formula is as follows:

[0063] ;

[0064] in, For a certain input feature or output feature, the first Data from each batch It is the minimum value of the data in the input or output feature. This represents the maximum value of the data in the input or output feature. for The result after normalization.

[0065] Combination Figure 4 The decoupling error neural network model structure in step 3 includes: a fully connected neural network (Linear) comprising one input layer, one output layer, and two hidden layers; the input layer has four neurons depending on the number of selected input features, and the output layer has two neurons depending on the number of selected output features; the Leaky ReLU activation function is used to perform nonlinear activation on each neuron layer; an additional BatchNorm layer and a Dropout layer are added between adjacent input layers, hidden layers, or output layers; the neural network parameter optimization method is Adaptive Moment Estimation (Adam), and the loss function is Mean Squared Error (MSE).

[0066] The formula for calculating the final value of the scaling factor after coupling the theoretical output of the decoupling algorithm in step 5 with the error prediction of the neural network is as follows:

[0067] ;

[0068] in, The theoretical actuator extension / retraction amount in the direction of nozzle pitch or yaw angle, calculated by the decoupling algorithm. To decouple the error neural network in The error prediction value of the actuator extension / retraction amount calculated by the decoupling algorithm in the corresponding direction. for The final value of the scaling factor after coupling the theoretical output of the decoupling algorithm in the corresponding direction with the error prediction of the neural network.

[0069] The coordinates of the follower point after the nozzle swing in step 1-3-2 are shown in the following formula:

[0070] ;

[0071] The coordinates of the follower point before its swing, which need to be calculated, are: The coordinates after the swing are ,in:

[0072] ;

[0073] in, For the nozzle axis and The angle between the planes, i.e., the pitch angle of the nozzle, and when the nozzle is facing... The value is positive in the positive direction of the axis. For the nozzle axis and The angle between the planes, i.e., the yaw angle of the nozzle, and when the nozzle is facing... The value is positive in the positive direction of the axis.

[0074] The system based on the above-mentioned decoupling control method for a swaying nozzle using a coupled neural network includes the following modules:

[0075] The data acquisition module is used to collect nozzle attitude data and actuator actual extension and retraction data, and to process the data using a decoupling algorithm to construct a nozzle attitude-decoupling error dataset. The dataset includes attitude data before and after nozzle swing, and actuator extension and retraction error data calculated by the decoupling algorithm when the attitude after nozzle swing is taken as input.

[0076] The data processing module is used to select the attitude data before and after nozzle swing and the actuator extension error data of the decoupling algorithm as the input and output features of the decoupling error neural network, and to normalize the feature data.

[0077] The decoupling error neural network model module is used to construct the decoupling error neural network model structure, determine the number of network layers and neurons, select the activation function, loss function and optimization algorithm type, and obtain the initial decoupling error neural network model.

[0078] The training module is used to divide the dataset into training set, validation set and test set, call optimization algorithm to automatically optimize the hyperparameters of the neural network model, verify the effectiveness of the model, save the model and related parameters, and obtain the trained decoupling error neural network model.

[0079] The extension / retraction command module is used to call the saved decoupling error neural network model, input the target attitude parameters of the nozzle into the neural network and the decoupling algorithm, perform forward propagation calculation of the neural network and solution of the decoupling algorithm, and couple the error prediction of the neural network with the theoretical output of the decoupling algorithm in real time to generate high-precision actuator extension / retraction commands, thereby achieving precise control of the nozzle attitude.

[0080] The accuracy of the decoupled control method and system was verified on a gyratory nozzle testing device. Yaw and pitch angle commands for the nozzle were randomly generated between -7° and 7°. The extension / retraction amounts of the two actuators were obtained through the decoupled control method and system and converted into actuation commands, which were then sent to the actuators for execution. The nozzle attitude was measured using an attitude sensor. This process was repeated 15 times, and the test results are as follows: Figures 6 to 8 As shown. By Figure 8 As can be seen from the dashed line, the errors of the nozzle yaw and pitch angles measured in 15 consecutive operations are all less than 0.1°, and the root mean square errors of the yaw and pitch angles are all less than 0.15°, indicating that the decoupling control method and system can accurately control the nozzle attitude.

[0081] Example

[0082] The steps of the decoupling algorithm in steps 1-3 are as follows:

[0083] Step 1-3-1: Establish a local coordinate system for the nozzle with the nozzle swing center as the origin, and determine the initial coordinates and initial distances of the connection points between the two actuators and the upper and lower nozzles.

[0084] Specifically, such as Figure 5 As shown, a local coordinate system is established fixed to the nozzle. The nozzle swing center position is Point, perpendicular to the engine centerline direction, , , , The upper and lower connection points of the two actuators are as follows: The direction is in the local coordinate system of the nozzle. In the positive direction of the axis, The direction is in the local coordinate system of the nozzle. Positive direction of the axis;

[0085] In the local coordinate system of the nozzle, the fixed point , and follower point , The initial coordinates are:

[0086]

[0087] in, , The distance from the upper end of each actuator to the center of the nozzle swing is respectively. Vertical and horizontal distances , The distance from the lower end of each actuator to the center of the nozzle swing is respectively. Vertical and horizontal distances;

[0088] in, and initial distance , and initial distance The distance can be calculated using the following formula:

[0089]

[0090] in, The distance between two points is given by the coordinates of the two points to be calculated. and ;

[0091] Step 1-3-2: Determine the transformation matrix of the coordinates under the rotational motion of the nozzle, and calculate the coordinates of the follower point at the connection point between the two actuators and the nozzle after the nozzle swings based on the coordinate transformation matrix;

[0092] Specifically: ignoring the axial rotation of the nozzle, when the nozzle rotates around the swing center... When the nozzle swings to a certain attitude, the local coordinate system fixed to the nozzle changes from... Transform to The corresponding coordinate transformation matrix is:

[0093]

[0094] in, For the nozzle axis and The angle between the planes, i.e., the pitch angle of the nozzle, and when the nozzle is facing... The value is positive in the positive direction of the axis. For the nozzle axis and The angle between the planes, i.e., the yaw angle of the nozzle, and when the nozzle is facing... The value is positive in the positive direction of the axis.

[0095] Follow-up point after nozzle oscillation , Transform into , The corresponding coordinates can be calculated using the following formula:

[0096]

[0097] The coordinates of the follower point before its swing, which need to be calculated, are: The coordinates after the swing are ;

[0098] Step 1-3-3: Calculate the distance between the two actuators and the upper and lower connection points of the nozzle after the nozzle swings, and the extension and retraction of the two actuators in the corresponding state, based on the transformed coordinates.

[0099] Specifically: the distance between the two actuators and the upper and lower connection points of the nozzle after the nozzle swings can be calculated using the distance formula from step 1-3-1 above. The calculation shows that the extension and retraction of the two actuators can be calculated using the following formula:

[0100]

[0101] in, Let be the initial distance between any actuator and the upper / lower connection point of the nozzle. This represents the distance between the actuator and the upper / lower connection point of the nozzle after the nozzle swings. This refers to the extension / retraction amount of the corresponding actuator. A positive value indicates that the actuator has extended. A negative value indicates that the actuator has contracted.

[0102] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0103] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for coupled neural network control of a wobble nozzle decoupling, characterized by, The method comprises the following steps: Step 1, collecting nozzle attitude data and actual actuator extension data, processing the data by using a decoupling algorithm, and constructing a nozzle attitude-decoupling error data set; the data set comprises attitude data before and after the nozzle swings, and actuator extension error data calculated by the decoupling algorithm with the nozzle attitude after the swing as input; Step 2, selecting the attitude data before and after the nozzle swings and the actuator extension error data of the decoupling algorithm as input features and output features of a decoupling error neural network respectively, and performing normalization processing on the feature data; Step 3, constructing a decoupling error neural network model structure, determining the number of network layers and the number of neurons, selecting an activation function, a loss function and an optimization algorithm type, and obtaining an initial decoupling error neural network model; Step 4, dividing the data set into a training set, a validation set and a test set, calling an optimization algorithm to automatically optimize the neural network model hyperparameters, performing model validity verification, saving the model and related parameters, and obtaining a trained decoupling error neural network model; Step 5, calling the saved decoupling error neural network model, inputting the target attitude parameters of the nozzle into the neural network and the decoupling algorithm, performing forward propagation calculation of the neural network and calculation of the decoupling algorithm, coupling the error prediction of the neural network and the theoretical output of the decoupling algorithm in real time, generating actuator extension instructions, and realizing control of the nozzle attitude.

2. The method of claim 1, wherein, The step 1 of constructing the nozzle attitude-decoupling error data set specifically comprises the following steps: Step 1-1, generating an extension instruction data set according to the extension range of the two actuators, randomly shuffling the sequence and converting it into an actuation instruction; Step 1-2, obtaining the actual extension amount of the actuator according to the internal code of the actuator; obtaining the attitude of the nozzle through an attitude sensor, including the pitch angle and the yaw angle of the nozzle; Step 1-3, inputting the obtained nozzle attitude into the decoupling algorithm to obtain the theoretical actuator extension amount under the corresponding attitude; Step 1-4, calculating the error of the decoupling algorithm according to the actual actuator extension amount obtained in the step 1-2 and the theoretical actuator extension amount in the step 1-3; Step 1-5, recording the obtained nozzle attitude, the nozzle attitude obtained in the previous collection, and the extension amount error of the decoupling algorithm, and if it is the first collection, only recording the measured nozzle attitude; Step 1-6, repeating the steps 1-2 to 1-5 until the collected data reaches a preset value, and constructing a nozzle attitude-decoupling error data set based on all the collected data.

3. The method of claim 2, wherein, The extension instruction data set in the step 1-1 is generated according to a uniform and random mixed generation strategy, specifically as follows: respectively in the extension and retraction range of the two actuators points, and combining the generated points of one actuator with the generated points of another actuator to obtain a set of actuator extension and retraction amount instructions; the random generation strategy is to respectively randomly select a generated point in the extension and retraction range of the two actuators, and combine the generated points of one actuator with the generated points of another actuator in sequence to obtain a set of actuator extension and retraction amount instructions; the uniform and random mixed generation strategy is to finally generate a set of actuator extension and retraction amount instructions by using uniform generation and random generation; wherein , , satisfies: .

4. The method of claim 2, wherein, The step 1-3 of inputting the nozzle attitude into the decoupling algorithm specifically comprises the following steps: Step 1-3-1, establishing a local coordinate system of the nozzle with the center of the nozzle swing as the origin, and determining the initial coordinates and initial distances of the two actuators and the upper and lower connection points of the nozzle; Step 1-3-2, determining the coordinate transformation matrix under the rotation motion of the nozzle, and calculating the coordinates of the follow-up points of the two actuators and the nozzle connection points according to the coordinate transformation matrix; Step 1-3-3, calculate the distance between the two actuators and the upper and lower connecting points of the nozzle after the nozzle swings according to the transformed coordinates, and the extension and contraction amounts of the two actuators in the corresponding state.

5. The method of claim 2, wherein, The decoupling algorithm error calculation formula of step 1-4 is: ； wherein, is the actual actuator extension in the pitch or yaw direction of the nozzle, is the is the theoretical actuator extension in the corresponding direction, is the is the error in the theoretical actuator extension in the corresponding direction calculated by the decoupling algorithm.

6. The method of claim 1, wherein, The input and output characteristics of the decoupling error neural network of step 2 are: The input characteristics are: ； wherein, is an input vector to a neural network, is a pitch angle feature measured by a pose sensor before the nozzle swing, is a yaw angle feature measured by a pose sensor before the nozzle swing, is a pitch angle feature measured by a pose sensor after the nozzle swing, is a yaw angle feature measured by a pose sensor after the nozzle swing; wherein the elements in each of the features satisfy: ； wherein, , are the pitch and yaw angles, respectively, of the first acquisition, the pitch and yaw angles, respectively, of the first acquisition before the nozzle oscillation, , are the pitch and yaw angles, respectively, of the first acquisition after the nozzle oscillation, wherein , , is the total number of acquisitions; The output characteristics are: ； wherein, is an output vector of the neural network, is an error characteristic of the theoretical actuator extension in the direction of the nozzle pitch angle calculated by the decoupling algorithm, is an error characteristic of the actuator extension in the direction of the nozzle yaw angle calculated by the decoupling algorithm.

7. The method of claim 1, wherein, The decoupling error neural network model structure of step 3 includes: using a fully connected neural network, containing 1 input layer, 1 output layer and 2 hidden layers; according to the number of selected input characteristics, the input layer is provided with 4 neurons, and according to the number of selected output characteristics, the output layer is provided with 2 neurons; using LeakyReLU activation function for nonlinear activation of each neuron layer; additionally adding a BatchNorm layer and a Dropout layer between adjacent input layer, hidden layer or output layer; the neural network parameter optimization method selects Adam, and the loss function selects MSE.

8. The method of claim 1, wherein, The final value calculation formula of the extension and contraction amount coupled after the theoretical output of the decoupling algorithm and the error prediction of the neural network of step 5 is: ； wherein, is the theoretical actuator extension in the direction of the nozzle pitch or yaw angle calculated by the decoupling algorithm, is the error prediction of the actuator extension in the direction calculated by the decoupling algorithm, is the error prediction of the actuator extension in the direction calculated by the decoupling algorithm, is the theoretical actuator extension in the direction of the nozzle pitch or yaw angle calculated by the decoupling algorithm, is the final value of the actuator extension in the direction calculated by the decoupling algorithm coupled with the error prediction of the neural network.

9. The method of claim 4, wherein, The coordinates of the follow-up points after the nozzle swings of step 1-3-2 are as follows: ； Wherein, the coordinates of the follow-up point before swinging are to be calculated , and the coordinates after swinging are , wherein: ； wherein is the angle between the nozzle axis and the plane of the axis, i.e. the pitch angle of the nozzle, and has a positive value when the nozzle is directed in the positive axis direction; is the angle between the nozzle axis and the plane of the axis, i.e. the yaw angle of the nozzle, and has a positive value when the nozzle is directed in the positive axis direction.

10. The system of coupled neural network's wobble nozzle decoupling control method according to any one of claims 1-9, wherein, It comprises the following modules: A data acquisition module is configured to acquire nozzle attitude data and actual actuator extension and contraction amount data, process the data using a decoupling algorithm, and construct a nozzle attitude-decoupling error data set; the data set comprises attitude data before and after the nozzle swings, and actuator extension and contraction amount error data calculated by the decoupling algorithm when the nozzle attitude after the nozzle swings is input; A data processing module is configured to select the attitude data before and after the nozzle swings and the actuator extension and contraction amount error data of the decoupling algorithm as input characteristics and output characteristics of a decoupling error neural network, and perform normalization processing on the characteristic data; A decoupling error neural network model module is configured to construct a decoupling error neural network model structure, determine the number of network layers and the number of neurons, select an activation function, a loss function and an optimization algorithm type, and obtain an initial decoupling error neural network model; A training module is configured to divide the data set into a training set, a validation set and a test set, call an optimization algorithm to automatically optimize neural network model hyperparameters, perform model validity verification, save the model and related parameters, and obtain a trained decoupling error neural network model; An extension and contraction amount instruction module is configured to call the saved decoupling error neural network model, input target attitude parameters of the nozzle into the neural network and the decoupling algorithm, perform forward propagation calculation of the neural network and calculation of the decoupling algorithm, real-time couple the error prediction of the neural network and the theoretical output of the decoupling algorithm, generate actuator extension and contraction amount instructions, and realize control of the nozzle attitude.

Images