Flexible docking method for rocket engine nozzle and combustion chamber

Abstract

The present invention provides a method for compliant docking of a rocket engine nozzle and a combustion chamber. The rocket engine nozzle is moved by a six-degree-of-freedom parallel platform along a preset motion path until it comes into contact with the rocket engine combustion chamber; the data returned by the six-axis force sensor, the position and attitude of the rocket engine nozzle are recorded, and the coordinates of the collision point are measured and recorded to form a training data set; the neural network model is trained with the training data set, and the prediction data set of the deep learning model is optimized by the wolf pack algorithm to improve the positioning accuracy of the collision prediction point; the force at the collision point is calculated based on the six-axis force sensor; according to the current force condition, the predicted force condition of the model, as well as the current position and attitude of the rocket engine nozzle and the collision point, the expected position and attitude of the rocket engine nozzle at the next time point are calculated through the compliant control mathematical model; the deep reinforcement learning algorithm is used to online learn the hyperparameters in the compliant control mathematical model to improve the docking effect.

Description

Technical Field

[0001] The present invention belongs to the field of rocket engine docking and general assembly, and specifically relates to a compliant docking method for a rocket engine combustion chamber and a rocket engine nozzle based on a six-axis force sensor and a six-degree-of-freedom parallel platform. Technical Background

[0002] The assembly of rocket engines is the most critical link in shortening the manufacturing cycle and ensuring product quality. During the assembly process, multi-layered stop redundancy limiting is used to improve the force-bearing characteristics and product sealing performance. The characteristics and difficulties of assembly quality assurance lie in the internal plug-in multi-step high-precision invisible assembly. When there is no contact, the engine nozzle can form a closed-loop control under the guidance of vision to meet the assembly requirements. However, when insertion and contact occur, the internal situation cannot be seen by vision. If only moving along the previously planned path at this time and not forming a closed loop with the vision system, it is very easy to cause bumps, shears, and extrusion deformations to the key components of the rocket engine during the docking process. The assembly deformation schematic diagram is as shown in Figures 2a to 2c shown, making the quality of the docking product unknown and posing extremely high safety hazards. The main difficulties are as follows: manufacturing errors and gravity deformation cause irregular changes in the reference relationship between multiple steps. When docking only relying on measuring the outermost reference, interference will occur in the internal precision mating steps. However, the internal interference situation is invisible and immeasurable in physical space, and the quality status cannot be accurately controlled.

[0003] The compliant control technology derived from the combination of a six-axis force sensor and a six-degree-of-freedom parallel platform provides a new means for the docking of the nozzle and combustion chamber of a rocket engine. The six-degree-of-freedom parallel platform is used to control the movement of the rocket engine nozzle along a preset movement path until it comes into contact with the rocket engine combustion chamber; repeat the collision test of the previous process, record the data returned by the six-axis force sensor, the position and attitude of the rocket engine nozzle, and measure and record the coordinates of the collision point to form a training data set; train a neural network model through the training data set, and optimize the prediction data set of the deep learning model through the wolf pack algorithm to improve the positioning accuracy of the collision prediction point; calculate the force condition at the collision point according to the six-axis force sensor; according to the current force condition, the predicted force condition of the model, and the current position and attitude of the rocket engine nozzle and the collision point, calculate the expected position and attitude of the rocket engine nozzle at the next time point through a compliant control mathematical model; during multiple collisions in a single docking process, use the deep reinforcement learning algorithm to online learn the hyperparameters in the compliant control mathematical model to improve the docking effect. Summary of the Invention

[0004] This method adds a deep learning model, a swarm intelligence algorithm, and a reinforcement learning model to the traditional compliant control concept, and optimizes all aspects of compliant control using information technology, artificial intelligence, etc., belonging to a brand-new technical method.

[0005] The technical solution adopted by the present invention to achieve the above object is as follows:

[0006] A compliant docking method for a rocket engine nozzle and a combustion chamber, comprising the following steps:

[0007] Step 1: Build a docking environment based on a six-degree-of-freedom parallel platform, install the main body of the rocket engine and the nozzle waiting for docking. In the non-collision stage, control the six-degree-of-freedom parallel platform to drive the nozzle to move along a pre-planned motion path until it reaches the position waiting for the collision test;

[0008] Step 2: In the collision test stage, according to the data sensed by the six-axis force sensor, calculate the range of the collision action point. Conduct multiple collision tests, record the training data set, build a deep learning model, predict the collision range, and use the wolf pack optimization algorithm to optimize the predicted collision range to obtain the final collision range;

[0009] Step 3: Predict the collision range for each collision point to obtain the docking line, and perform docking based on the docking line.

[0010] The said Step 2 includes the following steps:

[0011] Step 2.1: Conduct a random collision test. When the rocket engine nozzle collides with the main body of the rocket engine, immediately stop the movement of the six-degree-of-freedom parallel platform, and record the data collected by the six-axis force sensor at this time, the pose of the rocket engine nozzle, and the collision position range;

[0012] Step 2.2: According to the translational invariance of force, use the data of the six-axis force sensor to calculate the force direction at the collision position;

[0013] Step 2.3: Repeat Step 2.1 and Step 2.2 multiple times, record multiple groups of data, and form a training data set;

[0014] Step 2.4: Build a BP neural network model, with the data collected by the six-axis force sensor, the pose of the rocket engine nozzle, and the force direction at the collision position as the input, and the collision range of the nozzle as the output, and train the neural network model;

[0015] Step 2.5: Use the wolf pack algorithm to further optimize the prediction result set of the neural network model to improve the prediction accuracy of the collision point position of the rocket engine nozzle and narrow the prediction range.

[0016] The said six-axis force sensor is arranged between the six-degree-of-freedom parallel platform and the support tooling of the rocket engine nozzle, and is used to collect the force in three directions and the torque in three directions received by the sensor.

[0017] Calculate the force direction at the collision position through the following formula:

[0018] F x +F y +F z =F

[0019] Among them, F x represents the magnitude of the force in the X-axis direction, F y represents the magnitude of the force in the Y-axis direction, F z represents the magnitude of the force in the Z direction, and F represents the magnitude of the resultant force.

[0020] Step 2.4 includes the following steps:

[0021] Step 2.4.1: Construct a neural network model with two hidden layers;

[0022] Step 2.4.2: Construct a training dataset, where the input consists of three parts. The first part is the data collected by a six-axis force sensor; the second part is the pose data of the rocket engine nozzle during a collision; the third part is the direction of the force on the rocket engine nozzle during a collision. The output consists of two parts. The first part is the set of predicted positions acting on the rocket engine nozzle; the second part is the data of the calculated force magnitude at the collision point of the rocket engine nozzle;

[0023] Step 2.4.3: Construct a loss function to be optimized, the nth-order three-dimensional space mean square error function:

[0024]

[0025] Among them, x pi y pi z pi represent the predicted three-axis coordinate values, x i y i z i represent the actually measured three-axis coordinate values;

[0026] Step 2.4.4: Forward propagation: Calculate the theoretical output at the current moment according to the input data, neural network weight matrix, bias, and activation function;

[0027] Step 2.4.5: Backward propagation: Substitute the theoretical output calculated at the current moment and the output in the actual dataset into the loss function, and use the gradient descent method to optimize the loss function, and update the neural network weight matrix using the optimization result;

[0028] Step 2.4.6: Iteratively perform Step 2.4.4 and Step 2.4.5, continuously reduce the value of the loss function to optimize the neural network model until the prediction result output by the neural network model reaches the set prediction accuracy range.

[0029] Step 2.5 includes the following steps:

[0030] Step 2.5.1: Initialization: Set the scale of the artificial wolf pack as N. Each artificial wolf represents a possible collision point within the predicted collision range. The dimension of the search space is D. The spatial position of the i-th artificial wolf is represented as:

[0031] X i =(x i1 , x i2 , x i3 , …, x iD ), 1 ≤ i ≤ N, 1 ≤ d ≤ D

[0032] x id = x min + rand(0, 1)*(x max - x min )

[0033] Among them, x id represents the value of a possible collision point in a certain dimension. x min and x max represent the lower and upper limits of the value range respectively, that is, the lower and upper limits of the value of the possible collision point coordinates within the prediction range. rand(0, 1) represents a random number between 0 and 1;

[0034] Step 2.5.2: Competing for the leading wolf: Select q artificial wolves with the best fitness, that is, the q artificial wolves closest to the actual collision point among the randomly generated possible collision points. These q artificial wolves search in h directions around themselves. Let the current position of the competing wolf be P 0 , P 1 be the new position generated around the current position. If the newly generated position is closer to the actual collision point than the current position, then take the newly generated position as the current position and continue the search. When the wandering ends, take the competing wolf at the current optimal position as the leading wolf, and the remaining q - 1 wolves as exploring wolves. Then the position of the j-th point in the d-th dimension among the h points generated near the exploring wolf, γ jd , 1 ≤ j ≤ h can be expressed as:

[0035] γ jd = xx id + rand(0, 1)* step a

[0036] Among them, xx id is the current position of the i-th exploring wolf in the d-th dimension. step a refers to the wandering step size. rand(0, 1) represents a random number between 0 and 1;

[0037] Step 2.5.3: Alpha Wolf Summoning: The alpha wolf initiates a summoning behavior by howling, calling k scout wolves in the vicinity to converge towards the position of the alpha wolf. Then, the position of scout wolf i at the (t + 1)-th iteration in the d-dimensional variable space is expressed as:

[0038]

[0039] In the formula, represents the position of the alpha wolf in the d-dimensional space at the t-th generation, and step b represents the scout wolf's running step length, represents the value of this dimension at the current moment, represents the value of this dimension at the next moment;

[0040] Step 2.5.4: Surrounding the Prey: The scout wolves closer to the prey and the alpha wolf jointly surround the prey. The prey is the position of the alpha wolf, i.e., the optimal point found in the current state. The position closest to the prey is regarded as the moving position of the prey and is achieved through the following formula. For the wolf pack at the t-th generation, assume the position of the prey in the d-dimensional space is step c represents the attack step length. Then, the surrounding behavior of the wolf pack is expressed as:

[0041]

[0042] where rand(-1, 1) represents a random number between -1 and 1;

[0043] Step 2.5.5: Wolf Pack Update Mechanism: Sort the prey captured by the wolf pack according to the principle from strong to weak, that is, sort the existing collision points obtained from near to far from the true collision point.

[0044] Step 3 includes the following steps:

[0045] Step 3.1: Using the known force conditions, the position, velocity, and acceleration of the current collision point, calculate the kinematic information of the collision point at the next moment through the compliant control mathematical model:

[0046]

[0047] where M d 、D d 、K d represent the inertia characteristic, damping characteristic, and stiffness characteristic respectively, x d represent the acceleration, velocity, and position information to be calculated at the next moment respectively, x 0 represent the acceleration, velocity, and position information at the current moment respectively, F extRefers to the external force received;

[0048] Step 3.2: Use the method of reinforcement learning to train the learning parameters of agents M d , D d , K d during each collision;

[0049] Step 3.3: According to the trained neural network model, continue to determine the collision range at the next moment, and loop through Steps 2.5.2 to 3.2 to predict the collision range for each position on the docking path until the docking task is completed.

[0050] The said Step 3.2 includes the following steps:

[0051] Step 3.2.1: Establish the value space of inertia characteristics, damping characteristics, and stiffness characteristics according to the material, mass, and shape of the collision object;

[0052] Step 3.2.2: Regard the six-degree-of-freedom platform and the host computer controlling its movement as a whole agent, establish a reward and punishment mechanism for the agent, and for each collision situation, establish a reward and punishment system, that is, provide a reward or punishment index for the predicted position of each behavior made by the agent, so that the agent can learn three parameters;

[0053] Step 3.2.3: Establish a reward and punishment function, and maximize the reward and punishment function to achieve the learning effect of the agent.

[0054] The said reward and punishment function U t is:

[0055] U t = R t + γ * R t+1 + γ 2 * R t+2 + γ 3 * R t+3 + …

[0056] where R t is the quantization result of the reward or punishment at time t, and γ is the time decay factor.

[0057] The present invention has the following beneficial effects and advantages:

[0058] 1. Based on the deep learning model, the present invention fully learns data such as six-axis force sensors and six-degree-of-freedom platform poses, achieving the effect of high-precision prediction of the collision range.

[0059] 2. Based on the wolf pack algorithm, the present invention further improves the positioning accuracy of the collision range predicted by the deep learning model, reduces the prediction range, and improves the prediction effect.

[0060] 3. The present invention utilizes the mathematical model of compliant control, enabling the docking product to complete docking with high quality and efficiency solely with the assistance of a six-axis force sensor even when it is out of the visual guidance. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 is the flow chart of the compliant docking of the rocket engine;

[0062] Figure 2a is the deformation state of the multi-stage stepped shaft hole during the assembly process of the rocket engine Figure 1 ;

[0063] Figure 2b is the second figure of the deformation state of the multi-stage stepped shaft hole during the assembly process of the rocket engine;

[0064] Figure 2c is the deformation state of the multi-stage stepped shaft hole during the assembly process of the rocket engine Figure 3 ;

[0065] Figure 3 is the composition diagram of the physical device of the present invention;

[0066] Figure 4 is the flow chart of the wolf pack algorithm. DETAILED DESCRIPTION OF THE INVENTION

[0067] The present invention will be further described in detail below with reference to the drawings and embodiments.

[0068] The physical device of the present invention is composed as Figure 3 shown, mainly composed of a cabin position and attitude adjustment device and a six-axis force sensor. Each part of the physical device is a mature technology. Among them, the cabin position and attitude debugging device is a 6-degree-of-freedom parallel platform, generally known as the Stewart platform, which is a general device; the six-axis force sensor uses the six-axis force sensor produced by ATI Industrial Automation Co., Ltd. in the United States.

[0069] Step 1: Build a docking environment based on the six-degree-of-freedom parallel platform, install the main body of the rocket engine and the nozzle waiting for docking, and during the non-collision stage, control the parallel platform to drive the nozzle to move along the pre-planned movement path.

[0070] Step 2: In the collision test stage, according to the data sensed by the six-axis force sensor, calculate the range of the collision action point, and at the same time conduct multiple collision tests and record the data set to construct a deep learning model to improve the prediction accuracy of the collision range. On this basis, use the wolf pack optimization algorithm to optimize the predicted range to further improve the positioning accuracy.

[0071] Step 2.1: Conduct a random collision test. Once a collision occurs between the rocket engine nozzle and the main body of the rocket engine, immediately stop the movement of the six-degree-of-freedom parallel platform, and record the data collected by the six-axis force sensor at this time, the pose of the rocket engine nozzle, and the collision position range.

[0072] Step 2.2: According to the translational invariance of force, calculate the direction of the force at the collision position based on the data of the six-axis force sensor.

[0073] F x +F y +F z =F

[0074] Where F x represents the magnitude of the force in the X-axis direction, F y represents the magnitude of the force in the Y-axis direction, F z represents the magnitude of the force in the Z direction, and F represents the magnitude of the resultant force.

[0075] And record the data.

[0076] Table 1 Data Record (including the feedback data of the six-axis force sensor and the collision position)

[0077] Step 2.3: Repeat Step 2.1 and Step 2.2 multiple times, record multiple sets of data, and form a training data set.

[0078] Step 2.4: Build a BP neural network model. Use the data collected by the six-axis force sensor, the pose of the rocket engine nozzle, and the direction of the force generated by the collision as the input, and the collision range of the nozzle as the output. Train the neural network model. The specific method is as follows.

[0079] Step 2.4.1: Build a neural network model with two hidden layers

[0080] Step 2.4.2: Build a training data set. The input consists of three parts. The first part is the data collected by the six-axis force sensor. Each set of data is a six-dimensional vector [F x F y F z M Rx M Ry M Rz T (Where F x F y F z respectively represent the forces collected along the X, Y, and Z axes, and M Rx M Ry M Rz ​respectively representing the torques received in the directions around the X, Y, and Z axes); the second part is the pose data of the rocket engine nozzle during a collision, and each set of data is a six-dimensional vector [X Y Z R x R y R z T (where X, Y, and Z respectively represent the coordinates of the observation point of the nozzle on the X, Y, and Z axes in the absolute coordinate system, and R x R y R z represents the angles of deflection of the entire nozzle relative to the X, Y, and Z axes of the absolute coordinate system); the third part is the direction of the force on the rocket engine nozzle during a collision, and each set of data is a three-dimensional vector [x y z] T (where x, y, and z can form a unique vector in space); the output consists of two parts. The first part is the set of positions predicted to act on the rocket engine nozzle, and the data structure is a 3*n-dimensional matrix, indicating that the position set consists of n points, recording the coordinates of each point relative to the absolute coordinate system: The second part is the data of the magnitude of the force at the collision point of the rocket engine nozzle calculated, and each set of data is a three-dimensional vector [f x f y f z T .

[0081] Step 2.4.3: Construct the loss function to be optimized, the nth-order three-dimensional space mean square error function:

[0082]

[0083] where x pi y pi z pi represent the predicted three-axis coordinate values, and x i y i z i represent the actually measured three-axis coordinate values.

[0084] Step 2.4.4: Forward propagation, calculate the theoretical output at this time according to the current input, neural network weight matrix, bias, activation function, etc.

[0085] Step 2.4.5: Backward propagation, substitute the theoretical output calculated at this time and the output in the actual dataset into the loss function, and use the gradient descent method to optimize the loss function, and update the neural network weight matrix using the optimization result.

[0086] Step 2.4.6: Iteratively perform Step 2.4.4 and Step 2.4.5, continuously reduce the loss function, optimize the neural network model, until the acceptable prediction accuracy range is reached.​​

[0087] As Figure 4 shown, Step 2.5: Apply the wolf pack algorithm to further locate and optimize the prediction result set of the deep learning model, improve the prediction accuracy of the rocket engine nozzle collision point position, and narrow the prediction range.

[0088] Step 2.5.1: Initialization. Let the size of the artificial wolf pack be N. Each artificial wolf represents a possible collision point within the predicted collision range. The dimension of the search space is D (in this problem, since we hope to find the collision point in space, the dimension of the search space is three-dimensional). The spatial position of the i-th artificial wolf can be expressed as:

[0089] X i =(x i1 ,x i2 ,x i3 ,…,x iD ), 1 ≤ i ≤ N, 1 ≤ d ≤ D

[0090] x id =x min +rand(0,1)*(x max -x min )

[0091] where x id is the value of the possible collision point in a certain dimension, x min x max represent the lower and upper limits of the value range respectively, that is, the lower and upper limits of the possible collision point coordinates in the previous prediction range, and rand(0,1) represents a random number between 0 and 1.

[0092] Step 2.5.2: Compete for the lead wolf (i.e., the point with the best match between the solution computing power and the sensor collection among all the initialized artificial wolves). Select q artificial wolves with the best fitness (the position closest to the actual point) as competitors. These q artificial wolves search in h directions around themselves (the size of h is set artificially according to the accuracy requirement. The larger the value of h, the longer the search time, and vice versa). Let the current position of the competing wolf be P 0 ,P 1 is generated around the current position. If the newly generated position is better than the current position, then take the newly generated position as the current position and continue the search. When the wandering ends, take the competing wolf at the current optimal position as the lead wolf, and the remaining (q - 1) wolves as exploring wolves. Then the position of the j-th point in the d-th dimension among the h points generated near the exploring wolves, γ jd (1 ≤ j ≤ h) can be expressed as:

[0093] γ jd =xx id+rand(0,1)*step a

[0094] where xx id represents the current position of the i-th exploring wolf in the d-th dimension, and step a represents the wandering step size, and rand(0,1) represents a random number between 0 and 1.

[0095] Step 2.5.3: Alpha Wolf Summoning. The alpha wolf initiates a summoning behavior by howling to gather k exploring wolves around it to move closer to the position of the alpha wolf. Then, the position of the i-th exploring wolf in the d-th variable space at the (t + 1)-th iteration is expressed as:

[0096]

[0097] where represents the position of the t-th generation alpha wolf in the d-th dimension space, and step b represents the raiding step size of the exploring wolf, represents the value of this dimension at the current moment, represents the value of this dimension at the next moment.

[0098] Step 2.5.4: Sieging the Prey. The exploring wolves closer to the prey and the alpha wolf jointly siege the prey, and the position closest to the prey is regarded as the moving position of the prey. For the t-th generation wolf pack, assume the position of the prey in the d-th dimension space is step c represents the attack step size. Then, the sieging behavior of the wolf pack can be expressed as:

[0099]

[0100] where rand(-1,1) represents a random number between -1 and 1.

[0101] Through Step 2.5.4, the optimal point found after this round of iteration is obtained.

[0102] Step 2.5.5: Wolf Pack Update Mechanism. The prey captured by the wolf pack is distributed according to the principle of "from strong to weak", ultimately resulting in the weakest wolves starving to death and the stronger wolves being able to survive.

[0103] Step 3.1: Using the known force conditions, the position, velocity, and acceleration of the current collision point, calculate the kinematic information of the collision point at the next moment through the compliant control mathematical model:

[0104]

[0105] where M d , D d , K drespectively represent the inertial characteristic, damping characteristic and stiffness characteristic x d respectively represent the acceleration, velocity and position information at the next moment to be calculated x 0 respectively represent the acceleration, velocity and position information at the current moment

[0106] Step 3.2: In the mathematical model of compliant control, M d , D d , K d respectively represent the inertial characteristic, damping characteristic and stiffness characteristic. By adjusting these three parameters, the dynamic attributes of the controlled object can be changed to obtain better prediction results. The forms of the three parameters may be scalars, vectors, matrices or even functions. Different collision positions and different collision forms may cause the above three fixed parameters to fail to achieve the best prediction effect. Therefore, reinforcement learning is used to continuously train the agent to learn the parameters during each collision

[0107] Step 3.2.1: Provide the behavior set for the agent. From the perspective of this problem, the behavior set of the agent is the value range of the three parameters. The value spaces of the inertial characteristic, damping characteristic and stiffness characteristic are established according to the inherent physical characteristics of the colliding object materials, masses and shapes

[0108] Step 3.2.2: Provide the reward and punishment mechanism for the agent. For each collision situation, establish a clear reward and punishment system, that is, provide clear reward or punishment indicators for the position predicted by each behavior made by the agent, so that the agent can better learn the three parameters

[0109] Step 3.2.3: Establish a reward and punishment function and maximize the reward and punishment function to achieve the learning effect of the agent

[0110] U t =R t +γ*R t+1 +γ 2 *R t+2 +γ 3 *R t+3 +…

[0111] Step 3.3: According to the trained neural network model, continue to determine the collision range at the next moment, and repeatedly apply Steps 2.5 to 3.2 for each prediction until the docking task is completed

Claims

1. Method for compliant docking of rocket engine nozzle and combustion chamber, Characterized in that, It includes the following steps: Step 1: Build a docking environment based on a six-degree-of-freedom parallel platform, install the main body of the rocket engine and the nozzle waiting for docking. In the non-collision stage, control the six-degree-of-freedom parallel platform to drive the nozzle to move along the pre-planned motion path until it reaches the waiting collision test position; Step 2: In the collision test stage, according to the data sensed by the six-axis force sensor, calculate the range of the collision action point, conduct collision tests multiple times, record the training data set, build a deep learning model, predict the collision range, and use the wolf pack optimization algorithm to optimize the predicted collision range to obtain the final collision range; Step 3: Predict the collision range for each collision point to obtain the docking line, and perform docking based on the docking line; The said Step 2 includes the following steps: Step 2.1: Conduct a random collision test. When the rocket engine nozzle collides with the main body of the rocket engine, immediately stop the movement of the six-degree-of-freedom parallel platform, and record the data collected by the six-axis force sensor at this time, the pose of the rocket engine nozzle, and the collision position range; Step 2.2: According to the translational invariance of force, use the data of the six-axis force sensor to calculate the force direction at the collision position; Step 2.3: Repeat Step 2.1 and Step 2.2 multiple times, record multiple groups of data, and form a training data set; Step 2.4: Build a BP neural network model, using the data collected by the six-axis force sensor, the pose of the rocket engine nozzle, and the force direction at the collision position as inputs, and the collision range of the nozzle as the output, and train the neural network model; Step 2.5: Use the wolf pack algorithm to further optimize the positioning of the prediction result set of the neural network model to improve the prediction accuracy of the collision point position of the rocket engine nozzle and narrow the prediction range.

2. The method for compliant docking of rocket engine nozzle and combustion chamber according to claim 1, Characterized in that, The six-axis force sensor is arranged between the six-degree-of-freedom parallel platform and the support tooling of the rocket engine nozzle, and is used to collect the forces in three directions and the torques received around three directions by the sensor.

3. The method for compliant docking of rocket engine nozzle and combustion chamber according to claim 1, Characterized in that, The force direction at the collision position is calculated by the following formula: F x +F y +F z =F Among them, F x represents the magnitude of the force in the X-axis direction, and F y represents the magnitude of the force in the Y-axis direction, and F z represents the magnitude of the force in the Z direction, and F represents the magnitude of the resultant force.

4. The method for compliant docking of rocket engine nozzle and combustion chamber according to claim 1, Characterized in that, The said Step 2.4 includes the following steps: Step 2.4.1: Build a neural network model with two hidden layers; Step 2.4.2: Build a training data set, where the input consists of three parts. The first part is the data collected by the six-axis force sensor; the second part is the pose data of the rocket engine nozzle when a collision occurs; the third part is the force direction of the rocket engine nozzle when a collision occurs. The output consists of two parts. The first part is the set of positions predicted to act on the rocket engine nozzle; the second part is the data of the force magnitude at the collision point of the rocket engine nozzle calculated; Step 2.4.3: Construct the loss function to be optimized, the nth-order three-dimensional space mean square error function: Among them, x pi , y pi , z pi represent the predicted three-axis coordinate values, and x i , y i , z i represent the actually measured three-axis coordinate values; Step 2.4.4: Forward propagation: Calculate the theoretical output at the current moment based on the input data and the neural network weight matrix, bias and activation function; Step 2.4.5: Back propagation: Substitute the theoretical output calculated at the current moment and the output in the actual data set into the loss function, and use the gradient descent method to optimize the loss function, and use the optimization result to update the weight matrix of the neural network; Step 2.4.6: Iterate steps 2.4.4 and 2.4.5, continuously reducing the value of the loss function to optimize the neural network model until the prediction result output by the neural network model reaches the set prediction accuracy range.

5. The method for compliantly docking a rocket engine nozzle and a combustion chamber according to claim 1, It is characterized in that The step 2.5 comprises the following steps: Step 2.5.1: Initialization: Assume that the size of the artificial wolf pack is N, each artificial wolf represents a possible collision point within the predicted collision range, the dimension of the search space is D, and the spatial position of the i-th artificial wolf is expressed as: X i =(x i1 ,x i2 ,x i3 ,…,x iD ), 1 ≤ i ≤ N, 1 ≤ d ≤ D x id = x min + rand(0,1) * (x max - x min ) where x id represents the value of the possible collision point in a certain dimension, x min , x max respectively represent the lower and upper limits of the value range, that is, the lower and upper limits of the value of the possible collision point coordinates in the prediction range, and rand(0,1) represents a random number between 0 and 1; Step 2.5.2: Competing for the alpha wolf: Select q artificial wolves with the best fitness, that is, the q artificial wolves closest to the actual collision point among the randomly generated possible collision points. These q artificial wolves search in h directions around themselves. Let the current position of the competing wolf be P 0 , P 1 is the new position generated around the current position. If the newly generated position is closer to the actual collision point than the current position, then the newly generated position is taken as the current position and the search continues. When the wandering ends, the competing wolf at the current optimal position is taken as the alpha wolf, and the remaining q - 1 wolves are taken as scout wolves. Then, the position of the j-th point in the d-th dimension among the h points generated near the scout wolves is γ jd , 1 ≤ h can be expressed as: γ jd = xx id + rand(0,1)*tep a Among them, xx id is the current position of the $i$-th exploring wolf in the $d$-th dimension, and step a refers to the wandering step size, and rand(0,1) represents a random number between 0 and 1; Step 2.5.3: Alpha wolf calling: The alpha wolf initiates the calling behavior by howling, calling the surrounding k scout wolves to move closer to the alpha wolf. Then the position of scout wolf i in the t+1th iteration in the d-dimensional variable space is expressed as: In the formula, represents the position of the alpha wolf in the t-th generation in the d-th dimensional space, and step b represents the raiding step length of the exploring wolf, represents the value of this dimension at the current moment, represents the value of this dimension at the next moment; Step 2.5.4: Siege the prey: The scout wolves closer to the prey and the alpha wolf jointly siege the prey, where the prey is at the position of the alpha wolf, i.e., the optimal point found in the current state, and the position closest to the prey is regarded as the moving position of the prey. This is achieved through the following formula. For the t-th generation of wolf packs, let the position of the prey in the d-th dimensional space be step c represents the attack step length. Then the siege behavior of the wolf pack is expressed as: Among them, rand(-1,1 represents a random number between -1 and 1; Step 2.5.5: Wolf pack update mechanism: sort the prey captured by the wolf pack from strong to weak, that is, sort the existing collision points from near to far from the real collision point.

6. The method for compliantly docking a rocket engine nozzle and a combustion chamber according to claim 1, It is characterized in that The step 3 comprises the following steps: Step 3.1: Using the known force conditions, the position, velocity and acceleration of the current collision point, the kinematic information of the collision point at the next moment is calculated through the compliant control mathematical model: Among them, M d , D d , K d respectively represent the inertial characteristic, damping characteristic and stiffness characteristic, x d respectively represent the acceleration, velocity and position information at the next moment to be calculated, x 0 respectively represent the acceleration, velocity and position information at the current moment, and F ext refers to the external force received; Step 3.2: Train the learning parameters of M d , D d , K d agents using the method of reinforcement learning during each collision; Step 3.3: Based on the trained neural network model, continue to determine the collision range at the next moment, and repeat steps 2.5.2 to 3.2 in a loop to predict each collision range on the docking path until the docking task is completed.

7. The method for compliantly docking a rocket engine nozzle and a combustion chamber according to claim 6, It is characterized in that The step 3.2 comprises the following steps: Step 3.2.1: Establish the value space of inertia characteristics, damping characteristics and stiffness characteristics according to the material, mass and shape of the collision object; Step 3.2.2: Take the six-degree-of-freedom platform and the host computer that controls its motion as an intelligent agent, establish a reward and punishment mechanism for the intelligent agent, and establish a reward and punishment system for each collision situation, that is, provide reward or punishment indicators for the predicted position of each action made by the intelligent agent, so that the intelligent agent can learn three parameters; Step 3.2.3: Establish a reward and punishment function, and maximize the reward and punishment function to achieve the effect of intelligent agent learning.

Images