Intelligent cooperative control method for flexible lander of small celestial body

Abstract

The intelligent collaborative control method for a small celestial body flexible lander disclosed in the present invention belongs to the field of deep space exploration technology. The implementation method of the present invention is: using a flexible lander intelligent dynamics model based on deep neural network training instead of a discrete curvature dynamics model. Using an LSTM-based RNN to learn the sequence data of the discrete curvature dynamics model of the flexible lander and train the flexible lander intelligent dynamics model is beneficial to avoid the gradient vanishing problem caused by too long a sequence. At the same time, it can memorize the motion evolution law of the flexible lander, extract the motion state correlation relationship of the lander dynamics model data at the time level through the memory unit of the LSTM network, and use the motion state correlation relationship to improve the fitting efficiency of the flexible lander dynamics model. The present invention uses the flexible lander intelligent dynamics model to design the attachment control compensation term, reduce the lander motion error, and improve the autonomous control accuracy and real-time performance of the flexible lander during the attachment process.

Description

Technical Field

[0001] The present invention relates to an autonomous control method for a small celestial body during its attachment phase, and in particular to a flexible lander collaborative control method based on a recurrent neural network, belonging to the field of deep space exploration technology. Background Art

[0002] With the increasing frequency and intensiveness of small-body exploration activities, future small-body surface attachment missions will place even greater demands on precision and safety. Designing attachment schemes more suitable for small-body environments and developing smarter, more precise, and more risk-responsive lander control algorithms are key challenges. Due to the weak gravitational pull, numerous disturbances, and high uncertainty near small bodies, even minor control deviations can cause the lander to bounce and tumble upon contact, rendering traditional rigid lander designs and deterministic attachment control strategies unsuitable. Flexibility and intelligence are the future trends in small-body lander design. On the one hand, flexible structures can effectively absorb collision energy, increase the system's tolerance to control deviations, and reduce the probability of bouncing and tumbling. On the other hand, incorporating artificial intelligence into the lander's autonomous control algorithm can enhance the control system's adaptability to complex flexible structures and uncertain environments, reduce lander motion errors, and further improve attachment safety.

[0003] Currently, research on landing trajectory optimization and guidance methods for weak-gravity, multi-constraint environments is relatively mature. Using optimization methods based on optimal control or nonlinear parameter programming, rapid online generation of the lander's center-of-mass trajectory is now possible, fully accounting for requirements such as obstacle avoidance, control saturation, and fuel consumption reduction. However, the dynamic characteristics of flexible landers are more complex than those of traditional rigid landers. Flexible connection structures cause the motion of various internal components to be interconnected and coupled, resulting in deformation and deviations from the overall trajectory and attitude motion of the lander. This new characteristic places high demands on the design of autonomous control algorithms for attachment. Since the performance of control algorithms is closely related to the accuracy of the controlled object's dynamic modeling, establishing an accurate model for flexible landers that can be used online is key to improving control accuracy. Current modeling methods for flexible structures mainly include lumped-parameter linear simplified models and discrete curvature models based on finite element principles. The former is computationally efficient but has low accuracy, while the latter is highly accurate but computationally expensive, making it difficult to use online. In order to achieve high-precision modeling and online real-time application at the same time, it is necessary to introduce a machine learning algorithm to learn and fit the deviation between the above two models, so as to reduce the computational consumption of the precise dynamic model, thereby obtaining a high-computational efficiency and high-precision flexible lander attachment control algorithm and improving the attachment success rate. Summary of the Invention

[0004] The technical problem to be solved by the intelligent collaborative control method for a small celestial body flexible lander disclosed in the present invention is: using a recurrent neural network to learn and fit the high-precision discrete curvature dynamics model of the flexible lander to obtain an intelligent dynamics model of the flexible lander, and using the intelligent dynamics model of the flexible lander to design an attachment control compensation term to reduce the lander motion error and improve the autonomous control accuracy and real-time performance of the flexible lander during the attachment process.

[0005] The purpose of the present invention is achieved through the following technical solutions.

[0006] In response to the complex dynamic characteristics of the lander's flexible structure, large motion deviations, and the difficulty of traditional control algorithms to balance control accuracy and real-time computing, the present invention discloses an intelligent collaborative control method for a small celestial body flexible lander. This method introduces a recurrent neural network (RNN) based on a long short-term memory (LSTM) structure to learn and fit the flexible lander's dynamic model. The trained flexible lander intelligent dynamic model replaces the complex flexible lander discrete curvature dynamics model to output the internal forces between nodes of the flexible lander under different motion states, providing a real-time control instruction solution basis for the control system, improving the control accuracy of the attachment process, reducing the lander motion error, and enhancing the attachment safety. The specific implementation method is as follows: randomly giving the flexible lander an initial state, using a high-precision flexible lander discrete curvature dynamics model to perform recursive simulation of the small celestial body's flexible attachment trajectory, and obtaining a series of lander node motion states and flexible internal force data pairs under the flexible lander discrete curvature dynamics model. The node motion state and flexible internal force data pairs from the discrete curvature dynamics model of the flexible lander are substituted into the linear simplified dynamics model of the flexible lander. The corresponding internal forces are calculated using the linear simplified dynamics model and subtracted from the internal forces of the discrete curvature dynamics model to obtain node state and model internal force difference data pairs. Based on the sequential characteristics of these node state and model internal force difference data pairs, a recurrent neural network (RNN) based on a long short-term memory (LSTM) architecture is constructed. Supervised training is performed using the node state in each node state and model internal force difference data pair as input and the model internal force difference as output. During training, the selective update mechanism of the LSTM memory unit is utilized to avoid the vanishing gradient problem caused by excessively long sequences. The memory units are used to extract the temporal correlation between the motion states of the flexible lander dynamics model data, which is then used to improve the fitting efficiency of the flexible lander dynamics model. The training process is iterated until the generalization fitting error of the RNN fitting of the flexible lander dynamics model falls below a specified value, resulting in a trained flexible lander dynamics internal force difference model. The flexible lander's intelligent dynamics model is derived by superimposing its internal force difference model with its simplified linear model. This model replaces the discrete curvature dynamics model to improve the real-time performance of the control system's dynamic inversion. The internal force difference output by the internal force difference model in the intelligent dynamics model is used as an intelligent compensation term in the controller. This compensation term is used to compensate for the inverse calculation of the flexible lander control system's dynamics, improving the accuracy of the flexible lander's intelligent collaborative control.

[0007] The intelligent collaborative control method for a small celestial body flexible lander disclosed in the present invention comprises the following steps:

[0008] Step 1: Randomly give the initial state of the flexible lander, use a high-precision discrete curvature dynamics model of the flexible lander to perform recursive simulation of the flexible attachment trajectory of the small celestial body, and obtain a series of lander node motion states and flexible internal force data pairs under the discrete curvature dynamics model.

[0009] The implementation method of step one is:

[0010] The flexible lander for small celestial bodies consists of two parts. One part is N rigid modules for installing actuators and payloads, called nodes, and the other part is the flexible material that wraps and connects these N nodes. The flexible lander after unfolding is disc-shaped, and the nodes are distributed in a centrally symmetrical manner. The overall motion state of the lander is comprehensively reflected by the position, velocity, attitude and angular velocity information of the nodes. In the fixed coordinate system oxyz of the landing point of the target small celestial body, let the position, attitude quaternion, velocity and angular velocity vectors of the three nodes of the lander be r respectively. i ,q i 、v i and ω i (i=1,2,3,…,N), lander N in =13×N-dimensional state vector x is defined as

[0011]

[0012] The flexible lander discrete curvature dynamics model divides the flexible structure of the lander into non-overlapping triangular units based on the discrete curvature characteristics of the surface. The motion and deformation characteristics of the flexible lander are then split into the motion characteristics of each discrete unit. The interaction between the units is derived based on the mechanical properties of the flexible material, and a series of motion differential equations are established and solved. With the help of the flexible lander discrete curvature dynamics model, the flexible connection internal forces acting on the three nodes at the current moment are obtained. The dynamic equation for the change law of the reaction state vector x is expressed as

[0013]

[0014] Among them F c is the active control force acting on the three nodes, g is the surface gravity and spin inertia force of the small celestial body on the node, F fe and M fe are the flexible connection internal forces and internal moments of the nodes predicted by the discrete curvature dynamics model of the flexible lander, and F d and M d are the unmodeled internal forces, internal moments and random environmental disturbances. m is the equivalent mass of a single node, I is the node inertia matrix, and the symbol Represents direct multiplication of quaternions.

[0015] Use the above dynamic equation (2) to simulate the attachment of the flexible lander, recursively deduce the attachment trajectory, and give the initial state of the lander and the state of the attachment target.

[0016]

[0017] Define the total flight time of the attachment process as t f , active control force F c The position, velocity and acceleration of the attachment process are calculated using an open-loop guidance law based on a polynomial trajectory.

[0018]

[0019] Where the polynomial coefficient vector p 0i ~p 3i According to the conditions given in formula (3), we can get the expression of control force with respect to time t: c (t).

[0020] The above control sequence F c (t) is substituted into the dynamic equation (2) to perform the dynamic recursion of the attachment process and obtain an attachment trajectory under open-loop control. During the recursion process, the state of each time step and the flexible internal force data [x, F fe (x)] and record it. By randomly generating several different initial states within the specified range, perform attachment trajectory recursion on each of them to obtain enough simulation data and proceed to step 2.

[0021] Step 2: Substitute the node motion state and flexible internal force data under the flexible lander discrete curvature dynamics model obtained by simulation in step 1 into the flexible lander linear simplified dynamics model, use the flexible lander linear simplified dynamics model to calculate the corresponding internal force, and subtract it from the internal force of the discrete curvature dynamics model to obtain the node state and model internal force difference data pair.

[0022] The implementation method of step 2 is:

[0023] The flexible lander linear simplified dynamic model approximates the flexible connection between nodes to a six-degree-of-freedom spring-damper-torsion spring connection system, and the flexible internal forces and internal moments between nodes generated by the deformation of the flexible lander are equivalent to the spring-damper-torsion spring connection forces and moments. The flexible lander dynamic equation under the flexible lander linear simplified dynamic model is:

[0024]

[0025] The linear simplified dynamic model (5) of the flexible lander is similar to the discrete curvature dynamic model expression (2) of the flexible lander. The only difference is the different algorithms for the internal force and internal torque functions. Substitute the lander state data recorded in the simulation of step 1 into the internal force calculation function of the linear simplified dynamic model of the flexible lander to obtain the state-internal force data pair [x, F ls (x)], and then the two sets of data are subtracted to obtain the data pair of the lander state and the difference between the internal forces of the two models. As the training and testing datasets for neural network fitting, Go to step three.

[0026] Step 3: Based on the sequence characteristics of the node state and model internal force difference data pairs obtained in Step 2, a recurrent neural network (RNN) based on the long short-term memory (LSTM) structure is constructed. Supervised training is performed using the node states in the node state and model internal force difference data pairs as input and the model internal force difference as output. During training, the selective update mechanism of the LSTM structure's memory units is used to avoid the gradient vanishing problem caused by excessively long sequences. The memory units are used to extract the temporal relationship between the motion states of the flexible lander dynamics model data, and this relationship is used to improve the efficiency of fitting the flexible lander dynamics model. The training process is iterated until the generalization fitting error of the RNN fitting of the flexible lander dynamics model falls below a specified value, resulting in a trained flexible lander dynamics internal force difference model.

[0027] Step 3 is implemented as follows:

[0028] The state internal force difference data pair of a single simulation trajectory in step 2 is a set of data sequences, and the total flight time of the trajectory is t all , the simulation time step is h, then the sequence length is N s =t all / h+1, the sequence is as follows

[0029]

[0030] The input data of the recurrent neural network is the flexible lander state x, with dimension N in , the output is the model internal force difference Fitting The dimension is N out =3×N. Design the network hidden layer dimension to be N h The network structure consists of four layers in sequence: input normalization layer - ReLU fully connected layer - LSTM layer - fully connected regression layer. The forward propagation calculation process is as follows.

[0031] For the input sequence L in ={x k}(k=0,1,…,Ns ), after the input normalization layer, the sequence is in Dimension and x k The same calculation method is

[0032]

[0033] Sequence L1 is obtained by passing through the Relu fully connected layer The dimension is N h , calculated as

[0034]

[0035] The Relu function is defined as

[0036]

[0037] W L2 N h ×N in dimensional fully connected network coefficient matrix, b L2 N h dimensional bias vector.

[0038] Sequence L2 is passed through the LSTM layer to obtain the sequence The dimension is N h , calculated as

[0039]

[0040] where a k is the output unit, c k For memory units, is the replacement element of the memory unit, Γ u , Γ f , Γ o They are update gate, forget gate and output gate respectively, and the above variables are all N h dimensional vector, matrix W (c / u / f / o)(a / x) N is the N of each unit and gate h ×N h dimension coefficient matrix, b c / u / f / o N h dimensional bias vector. The symbol * represents the multiplication of vector elements. The sigmoid function is defined as follows

[0041]

[0042] In addition, when k=1, a k-1 and c k-1 All zero vectors.

[0043] Sequence L3 passes through the fully connected regression layer to obtain the final output sequence The dimension is N out , calculated as

[0044]

[0045] W L4 N out ×N h dimensional fully connected network coefficient matrix, b L4 N out dimensional bias vector.

[0046] This recurrent network calculates the mean squared error between the output layer and the training data as the regression loss function for backpropagation. During actual training, the test mean squared error (generalization error) on the test set data, which is less than a specified value, is considered a sign of training effectiveness. After training is complete, proceed to step 4.

[0047] Step 4: Superimpose the flexible lander dynamic internal force difference model trained in Step 3 with the linear simplified model to obtain the flexible lander intelligent dynamic model. Use this flexible lander intelligent dynamic model instead of the discrete curvature dynamic model to improve the real-time performance of the control system dynamic inversion. Use the internal force difference output by the internal force difference model in the flexible lander intelligent dynamic model as the controller's intelligent compensation term. This compensation term is used to compensate for the flexible lander control system dynamic inversion calculation, improving the accuracy of the flexible lander's intelligent collaborative control.

[0048] Step 4 is implemented as follows:

[0049] The flexible lander dynamic internal force difference model trained in step 3 is superimposed on the flexible lander linear simplified model of formula (5) to obtain the flexible lander intelligent dynamic model, which replaces the flexible lander discrete curvature dynamic model of formula (2). The node center of mass acceleration formula of the intelligent dynamic model becomes as follows:

[0050]

[0051] Take the polynomial guidance law designed in step 1 as the nominal guidance law, and define the nominal guidance acceleration, nominal velocity, and nominal position vector as

[0052]

[0053] The flexible attachment terminal sliding mode TSM control law is designed using the nominal guidance law and the intelligent dynamics model, and the sliding mode error vector is defined as

[0054]

[0055] The sliding surface is defined as

[0056]

[0057] Wherein β, p, and q are positive integers, p>q, and 1<p / q<2.

[0058] The control acceleration formula of the flexible lander intelligent cooperative control law designed based on the Lyapunov stability principle is:

[0059]

[0060] in One item is the controller compensation item provided by the flexible lander intelligent dynamics model.

[0061] Through the compensation Compensate for the dynamic inversion calculation of the flexible lander control system to improve the accuracy of the flexible lander's intelligent collaborative control.

[0062] Beneficial effects:

[0063] 1. The intelligent collaborative control method for a small celestial body flexible lander disclosed in the present invention uses a flexible lander linear simplified model superimposed on a flexible lander intelligent dynamics model to replace the flexible lander discrete curvature dynamics model. This not only retains the high precision advantage of the discrete curvature dynamics model, but also significantly improves the efficiency of flexible lander dynamics prediction, making the present invention have the characteristics of fast calculation speed and high precision. It can effectively solve the problem that the high-precision discrete curvature dynamics model of flexible lander dynamics has a large amount of calculation and cannot be applied online in real time, thereby improving the control accuracy during the lander attachment process.

[0064] 2. The intelligent collaborative control method for a small celestial body flexible lander disclosed in the present invention uses a recurrent neural network (RNN) based on a long short-term memory (LSTM) model to learn and fit the dynamic recursive sequence data of the flexible lander's attachment simulation trajectory. It can efficiently discover and memorize the law of the time evolution of the motion state of each node of the lander during the attachment process, and can avoid the gradient vanishing problem caused by an excessively long sequence, thereby improving the fitting accuracy of the flexible lander's intelligent dynamic model to the discrete curvature dynamic model. In addition, the motion state correlation relationship of the flexible lander's dynamic model data at the time level is extracted through the memory unit of the LSTM network, and the motion state correlation relationship is used to improve the fitting efficiency of the flexible lander's dynamic model.

[0065] 3. In the intelligent collaborative control method for a small celestial body flexible lander disclosed in the present invention, the RNN does not directly fit the discrete curvature dynamics model, but instead fits the difference between the discrete curvature dynamics model and the linear simplified dynamics model on the basis of superimposing the linear simplified dynamics model, thereby reducing the training difficulty of the RNN and improving the RNN's prediction accuracy for the flexible lander dynamics model.

[0066] 4. The intelligent collaborative control method for the small celestial body flexible lander disclosed in the present invention combines the intelligent dynamics model with the polynomial guidance theory and the terminal sliding mode control method, and has the characteristics of high control accuracy and strong system stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 This is a flowchart of the steps of the intelligent collaborative control method for a small celestial body flexible lander disclosed in the present invention;

[0068] Figure 2 Schematic diagram of the linear simplified dynamic model and discrete curvature dynamic model of the flexible lander;

[0069] Figure 3 Generating a three-dimensional image of a flexible lander attachment trajectory cluster using a flexible lander discrete curvature dynamics model in Example 1;

[0070] Figure 4 : is a comparison diagram of the flexible internal force of the discrete curvature dynamic model and the linear simplified dynamic model of a certain simulation trajectory in Example 1, wherein: Figure 4 (a) is the flexible internal force output by the discrete curvature dynamics model of the flexible lander, Figure 4 (b) is the flexible internal force output by the linear simplified dynamic model of the flexible lander. Figure 4 (c) is the difference between the flexible internal forces of the above two models;

[0071] Figure 5 Fitting the mean square error change curve of the recurrent neural network training and testing process in Example 1;

[0072] Figure 6 Comparison curves of trajectory node position errors for collaborative attachment control using the flexible lander linear simplified dynamics model and the intelligent dynamics model in Example 1.

[0073] Figure 7 Comparison curves of trajectory node velocity errors for collaborative attachment control using the flexible lander linear simplified dynamics model and the intelligent dynamics model in Example 1.

[0074] Figure 8 Comparison curves of trajectory node control forces for collaborative adhesion control using the flexible lander linear simplified dynamics model and the intelligent dynamics model in Example 1.

[0075] Figure 9 Comparison curves of trajectory node flexibility internal forces for collaborative adhesion control using the flexible lander linear simplified dynamics model and the intelligent dynamics model in Example 1. DETAILED DESCRIPTION

[0076] In order to better illustrate the purpose and advantages of the present invention, the invention is further described below with reference to the accompanying drawings and examples.

[0077] Example 1:

[0078] In order to verify the feasibility of the method, a small celestial body attachment mission was taken as an example to carry out flexible lander intelligent dynamic model fitting and intelligent collaborative control simulation calculation.

[0079] like Figure 1 As shown, the convex curvature landing trajectory fuel consumption optimization design method disclosed in this embodiment is specifically implemented in the following steps:

[0080] Step 1: Randomly give the initial state of the flexible lander, use a high-precision discrete curvature dynamics model of the flexible lander to perform recursive simulation of the flexible attachment trajectory of the small celestial body, and obtain a series of lander node motion states and flexible internal force data pairs under the discrete curvature dynamics model.

[0081] The simulation uses a disk-shaped flexible lander with N = 3 rigid nodes, which are symmetrically distributed around the center of the lander. The relative distance between the nodes in the original state is 1.04m. The schematic diagram of the flexible lander, the schematic diagram of the linear simplified dynamic model of the flexible lander, and the schematic diagram of the discrete curvature dynamic model are shown in Figure 2. Figure 2 As shown in the discrete curvature dynamics model, the node mass is 150 kg, the flexible material is a typical silicone rubber material, and the Young's modulus is given as 1×10 6 , Poisson's ratio 0.4, density 1kg / m 3 ; Equivalent to the linear simplified dynamic model, since the mass of the flexible material is not considered, the equivalent mass of the node increases to 333.5kg, the spring stiffness is 1708.4N / m, the damping coefficient is 2.6Ns / m, the torsional stiffness of the torsion spring in the node system in the x-axis direction is 29.3Nm / rad, the torsional stiffness in the y-axis direction is 13.9Nm / rad, and the torsional damping coefficient in both directions is 1.8Nms / rad. In the target small body landing point coordinate system, the initial center of mass position of the flexible lander is randomly specified within the rectangular range of [±30±30 50±10]m, and the initial center of mass velocity is randomly specified within the range of [±0.1±0.1±0.1]m / s. The target attachment position is the origin of the coordinate system, and the target velocity is zero. The total flight time is set to 50s, and the dynamic recursive time step is 0.1s. During the attachment process, each node is subject to an upper limit of 0.1m / s. 2 The random perturbation acceleration is recursively deduced for a total of 5000 trajectories, and the state internal force data are recorded. The three-dimensional curves of some of the trajectories are as follows Figure 3 As shown in the figure, due to the influence of flexible internal force and disturbance force on the node, the terminal attachment error is significant when only the open-loop guidance law is used.

[0082] Step 2: Substitute the node motion state and flexible internal force data under the flexible lander discrete curvature dynamics model obtained by simulation in step 1 into the flexible lander linear simplified dynamics model, use the flexible lander linear simplified dynamics model to calculate the corresponding internal force, and subtract it from the internal force of the discrete curvature dynamics model to obtain the node state and model internal force difference data pair.

[0083] Randomly select one of the 5000 attachment trajectories derived in step 1 and compare the internal force calculations of the two dynamic models. The comparison is as follows: Figure 4 As shown, the three-node internal force data calculated by the two models are of the same magnitude and exhibit similar overall trends. However, due to the more sophisticated modeling, the internal force curve of the discrete curvature dynamics model reflects higher-dimensional modal information and exhibits a higher frequency of periodic changes compared to the linear simplified dynamics model. This is more evident in the curve showing the difference in internal force calculations between the two models.

[0084] Step 3. Based on the sequence characteristics of the node state and model internal force difference data pairs obtained in Step 2, a recurrent neural network (RNN) based on the long short-term memory structure (LSTM) is constructed. The node state in the node state and model internal force difference data pairs is used as input, and the model internal force difference is used as output for supervised training. During training, the selective update mechanism of the memory unit of the LSTM structure is used to avoid the gradient vanishing problem caused by excessively long sequences. The memory unit is used to refine the motion state correlation relationship of the flexible lander dynamics model data at the temporal level. The motion state correlation relationship is used to improve the fitting efficiency of the flexible lander dynamics model. The training process is iterated until the generalization fitting error of the RNN fitting of the flexible lander dynamics model is lower than the specified value, and the trained flexible lander dynamics internal force difference model is obtained.

[0085] The training data extracted in step 2 is used for mean square error regression training, and the number of iterations is set to 3000. The output mean square error (RMSE) of the test data is less than 0.5 as the condition for judging whether the training accuracy meets the standard. The training process is as follows: Figure 5 As shown in the figure, as the training progresses, the mean square error and comprehensive loss function of the training set and test set show a significant downward trend. At the end of the training, the RMSE of the test set is about 0.46, achieving the expected training effect.

[0086] Step 4: Superimpose the flexible lander dynamic internal force difference model trained in Step 3 with the linear simplified model to obtain the flexible lander intelligent dynamic model. Use this flexible lander intelligent dynamic model instead of the discrete curvature dynamic model to improve the real-time performance of the control system dynamic inversion. Use the internal force difference output by the internal force difference model in the flexible lander intelligent dynamic model as the controller's intelligent compensation term. This compensation term is used to compensate for the flexible lander control system dynamic inversion calculation, improving the accuracy of the flexible lander's intelligent collaborative control.

[0087] The flexible lander intelligent dynamics model and the linear simplified dynamics model are used to provide node state prediction for the terminal sliding mode controller. A comparative simulation of flexible adhesion cooperative control is performed under the same simulation conditions as step 2. The error between the adhesion trajectory and the nominal polynomial trajectory it tracks is plotted. The position error curve is shown in Figure 6 As shown in the figure, the trajectory position error is smaller when the controller uses intelligent compensation, but it increases significantly when only the linear simplified dynamic model is used.

[0088] The speed error curve is as follows Figure 7 As shown in the figure, the trajectory velocity error is smaller when the controller uses intelligent compensation, but it increases significantly when only the linear simplified dynamic model is used.

[0089] The node control force curve is as follows Figure 8 As shown in the figure, the oscillation amplitude of the trajectory control force is small when the controller uses intelligent compensation, and the oscillation amplitude of the control force increases significantly when only the linear simplified dynamic model is used.

[0090] The node flexible inner curve is as follows Figure 9 As shown in the figure, the amplitude of the flexible internal force is small when the controller uses intelligent compensation, reflecting the good stability of the lander. When only the linear simplified dynamic model is used, the amplitude of the flexible internal force increases significantly.

[0091] Based on the above simulation results, it is found that the trained flexible lander intelligent dynamics model can significantly improve the performance of the feedback controller when the flexible lander is attached to the surface of a small celestial body. The intelligent collaborative control method established on this basis can effectively improve the control accuracy of the flexible lander and thus improve the attachment safety.

[0092] The above specific description further illustrates the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent collaborative control method for a small celestial body flexible lander, characterized by: The following steps are included: Step 1: Randomly assign the initial state of the flexible lander and use a high-precision discrete curvature dynamics model of the flexible lander to perform recursive simulation of the flexible attachment trajectory of the small celestial body, obtaining a series of lander node motion states and flexible internal force data pairs under the discrete curvature dynamics model; Step 2: Substitute the node motion state and flexible internal force data pairs under the flexible lander discrete curvature dynamics model obtained by simulation in step 1 into the flexible lander linear simplified dynamics model, calculate the corresponding internal forces using the flexible lander linear simplified dynamics model, and subtract them from the internal forces of the discrete curvature dynamics model to obtain node state and model internal force difference data pairs; Step 3: Based on the sequence characteristics of the node state and model internal force difference data pairs obtained in step 2, a recurrent neural network (RNN) based on the long short-term memory (LSTM) structure is constructed. The node state in the node state and model internal force difference data pairs is used as input, and the model internal force difference is used as output for supervised training. During training, the selective update mechanism of the memory unit of the LSTM structure is used to avoid the gradient vanishing problem caused by overly long sequences. The memory unit is used to extract the motion state correlation relationship of the flexible lander dynamics model data at the time level, and the motion state correlation relationship is used to improve the fitting efficiency of the flexible lander dynamics model. The training process is iterated until the generalization fitting error of the RNN fitting of the flexible lander dynamics model is lower than the specified value, and the trained flexible lander dynamics internal force difference model is obtained. Step 4. Superimpose the flexible lander dynamic internal force difference model obtained by training in step 3 with the flexible lander linear simplified dynamic model to obtain the flexible lander intelligent dynamic model. Use the flexible lander intelligent dynamic model instead of the discrete curvature dynamic model to improve the real-time performance of the control system dynamic inversion. Use the internal force difference output by the internal force difference model in the flexible lander intelligent dynamic model as the controller intelligent compensation item. Use the compensation item to compensate the dynamic inversion calculation of the flexible lander control system to improve the intelligent collaborative control accuracy of the flexible lander.

2. The intelligent collaborative control method for a small celestial body flexible lander according to claim 1, characterized in that: The implementation method of step one is: The flexible lander for small celestial bodies consists of two parts: one is N rigid modules for installing actuators and payloads, called nodes, and the other is flexible materials that wrap and connect these N nodes. After unfolding, the flexible lander is disc-shaped, and the nodes are distributed in a centrally symmetrical manner. The overall motion state of the lander is comprehensively reflected by the position, velocity, attitude and angular velocity information of the nodes. In the fixed coordinate system oxyz of the landing point of the target small celestial body, the position, attitude quaternion, velocity and angular velocity vectors of the three nodes of the lander are r i ,q i 、v i and ω i (i=1,2,3,…,N), lander N in =13×N-dimensional state vector x is defined as The discrete curvature dynamics model for flexible landers divides the flexible structure of the lander into non-overlapping triangular units based on the discrete curvature characteristics of the surface. The motion and deformation characteristics of the flexible lander are then broken down into the kinematic characteristics of each discrete unit. The interactions between the units are deduced based on the mechanical properties of the flexible material, and a series of differential equations of motion are established and solved. With the help of the flexible lander discrete curvature dynamics model, the flexible connection internal forces acting on the three nodes at the current moment are obtained. The expression of the flexible lander discrete curvature dynamics model of the reaction state vector x is: Among them F c is the active control force acting on the three nodes, g is the surface gravity and spin inertia force of the small celestial body on the node, F fe and M fe are the flexible connection internal forces and internal moments of the nodes predicted by the discrete curvature dynamics model of the flexible lander, and F d and M d They are the unmodeled internal forces, internal moments and random environmental disturbances; m is the equivalent mass of a single node, I is the node inertia matrix, and the symbol Represents direct multiplication of quaternions; Use the above dynamic model expression (2) to simulate the attachment of the flexible lander, recursively deduce the attachment trajectory, and give the initial state of the lander and the attachment target state Define the total flight time of the attachment process as t f , active control force F c Calculated using an open-loop guidance law based on polynomial trajectories; for node i, the acceleration, velocity, and position during the attachment process are as follows Where the polynomial coefficient vector p 0i ~p 3i According to the conditions given in formula (3), we can get the expression of control force with respect to time t: c (t); The above expression F c (t) is substituted into the dynamic equation (2) to perform the dynamic recursion of the attachment process and obtain an attachment trajectory under open-loop control. During the recursion process, the state of each time step and the flexible internal force data given by the discrete curvature dynamic model of the corresponding flexible lander are combined. Record it; randomly generate several different initial states within the specified range, perform attachment trajectory recursion to obtain enough simulation data, and then proceed to step 2.

3. The intelligent collaborative control method for a small celestial body flexible lander according to claim 2, characterized in that: The implementation method of step 2 is: The flexible lander linear simplified dynamic model approximates the flexible connection between nodes to a six-degree-of-freedom spring-damper-torsion spring connection system, and the flexible internal force and internal torque between nodes generated by the deformation of the flexible lander are equivalent to the spring-damper-torsion spring connection force and torque; the flexible lander dynamic equation under the flexible lander linear simplified dynamic model is: The linear simplified dynamic model of the flexible lander (5) is similar to the discrete curvature dynamic model expression of the flexible lander (2). The only difference is the different algorithms for the internal force and internal torque functions. Substitute the lander state data recorded in the simulation of step 1 into the internal force calculation function of the linear simplified dynamic model of the flexible lander to obtain the state-internal force data pair under the linear simplified dynamic model of the flexible lander. Then, the two sets of data are subtracted to obtain the data pair of the lander state and the difference between the internal forces of the two models. As the training and test data sets for neural network fitting, where ▽F in (x) = F fe (x)-F ls (x), go to step 3.

4. The intelligent collaborative control method for a small celestial body flexible lander according to claim 3, characterized in that: The implementation method of step three is: The data pairs of the flexible lander state and the difference between the internal forces of the two models for a single simulation trajectory in step 2 are taken as a set of data sequences, and the total flight time of the trajectory is recorded as t all , the simulation time step is h, then the sequence length is N s =t all / h+1, the sequence is as follows The input data of the recurrent neural network is the flexible lander state x, with dimension N in , the output is the model internal force difference ▽F in Fitting of (x) The dimension is N out =3×N; the hidden layer dimension of the network is designed to be N h ,The network structure is a four-layer sequence: input normalization layer—Relu fully connected layer—LSTM layer—fully connected regression layer; The forward propagation calculation process is as follows; For the input sequence L in ={x k }(k=0,1,…,N s ), after the input normalization layer, the sequence is in Dimension and x k The same calculation method is Sequence L1 is obtained by passing through the Relu fully connected layer The dimension is N h , calculated as The Relu function is defined as W L2 N h ×N in dimensional fully connected network coefficient matrix, b L2 N h dimensional bias vector; Sequence L2 is passed through the LSTM layer to obtain the sequence The dimension is N h , calculated as where a k is the output unit, c k For memory units, is the replacement element of the memory unit, Γ u , Γ f , Γ o They are update gate, forget gate and output gate respectively, and the above variables are all N h dimensional vector, matrix W (c / u / f / o)(a / x) N is the N of each unit and gate h ×N h dimension coefficient matrix, b c / u / f / o N h dimensional bias vector; the symbol * represents the multiplication of vector elements; the sigmoid function is defined as follows In addition, when k=1, a k-1 and c k-1 All are zero vectors; Sequence L3 passes through the fully connected regression layer to obtain the final output sequence The dimension is N out , calculated as W L4 N out ×N h dimensional fully connected network coefficient matrix, b L4 N out dimensional bias vector; This recurrent neural network calculates the mean square error between the output layer and the training data as the regression loss function for back propagation. During the actual training process, the test mean square error of the test set data, that is, the generalization error is less than the specified value, which is used as a sign that the training effect has achieved the expected result. After the training is completed, proceed to step 4.

5. The intelligent collaborative control method for a small celestial body flexible lander according to claim 4, characterized in that: The implementation method of step 4 is: The flexible lander dynamic internal force difference model trained in step 3 is superimposed on the flexible lander linear simplified dynamic model of formula (5) to obtain the flexible lander intelligent dynamic model, which replaces the flexible lander discrete curvature dynamic model of formula (2). The node center of mass acceleration formula of the intelligent dynamic model becomes as follows: Take the polynomial guidance law designed in step 1 as the nominal guidance law, and define the nominal guidance acceleration, nominal velocity, and nominal position vector as The flexible attachment terminal sliding mode TSM control law is designed using the nominal guidance law and the intelligent dynamics model, and the sliding mode error vector is defined as The sliding surface is defined as Where β, p, and q are positive integers, p>q, and 1<p / q<2; The control acceleration formula of the flexible lander intelligent cooperative control law designed based on the Lyapunov stability principle is: in One is the controller compensation term provided by the flexible lander intelligent dynamics model; Through the compensation Compensate for the dynamic inversion calculation of the flexible lander control system to improve the accuracy of the flexible lander's intelligent collaborative control.

Images