Satellite electromagnetic docking process generalization simulation method based on FCNN

Abstract

The application discloses a satellite electromagnetic docking process generalization simulation method based on FCNN and belongs to the field of spacecraft electromagnetic docking dynamics simulation, and comprises the following steps: firstly, magnetic field data of different types of electromagnetic actuators under random excitation current is generated through finite element simulation, an FCNN-D model is constructed and trained, and prediction from actuator parameters, spatial position and current input to magnetic field strength is realized; secondly, force data of the actuator in an external magnetic field is collected, an FCNN-F model is constructed to predict electromagnetic force / torque with the actuator type, pose and current as input; subsequently, dynamics simulation is carried out, the force / torque calculated by the FCNN-F is input step by step, electromagnetic force is calculated according to the current pose in each simulation, and the satellite pose is updated until the boundary condition is met; finally, the docking process is visualized. The application utilizes FCNN to avoid the electromagnetic force / torque model, solves the defect that the application scale of electromagnetic simulation is limited, and obtains an electromagnetic docking simulation method with "full scale" and "high precision".

Description

TECHNICAL FIELD

[0001] The present application relates to the field of spacecraft electromagnetic docking dynamics simulation, in particular to a method for simulating the interaction of electromagnetic docking actuators and the motion process of spacecraft electromagnetic docking based on a full connection neural network. BACKGROUND

[0002] With the development of micro-nano satellites and space stations, spacecraft in-orbit interaction is becoming more and more frequent, and the traditional jet propulsion docking technology has problems such as fuel consumption and plume impact, which has a negligible impact on the service life of spacecraft. Electromagnetic docking meets the current docking high frequency, small amplitude in-orbit maneuvering demand with its fuel consumption, low docking impact and other advantages. The main problem faced by electromagnetic docking is the strong nonlinear characteristics of electromagnetism and the coupling characteristics of electromagnetism and kinematics. However, the Biot-Savart law in electromagnetism does not have an analytical solution in engineering, and most of the various magnetic dipole models currently used for theoretical analysis are only applicable at a specific scale, which cannot meet the precision requirements of small-scale docking processes. In addition, most simulation analyses in related fields currently separate the electromagnetic finite element simulation analysis of electromagnetic actuators and the dynamics and control system simulation analysis of micro-nano satellites in-orbit motion, the former is mostly limited to simple motion in a static field, and the latter mostly uses a simple model obtained by linearization approximation, which is difficult to adapt to the high coupling characteristics of electromagnetism and kinematics of electromagnetic actuators. Therefore, it is necessary to seek a complex dynamic model of electromagnetism and kinematics coupling through data-driven and numerical analysis methods, which is applicable to all motion scales.

[0003] Most simulation analyses in related fields currently separate the electromagnetic actuator type and distribution design of micro-nano satellites and the interaction between electromagnetic actuators on the tracking star and the target star, making it difficult to combine and unify the two. The current single-star research process in the aerospace industry is difficult to meet the demand for space resource competition with the construction of low-orbit micro-nano satellite constellations. Therefore, it is particularly necessary to develop a general-purpose high-integration theoretical research platform for micro-nano satellites containing electromagnetic actuators to simulate the docking process of different types of satellites under different working conditions, in order to improve the efficiency of series small satellite research and management. SUMMARY

[0004] In view of the limitations of current electromagnetic simulation methods, the present application proposes a generalization simulation method for satellite electromagnetic docking process based on FCNN, aiming to establish a multi-dimensional simulation method for the electromagnetic docking process of a CubeSat with four electromagnetic docking actuators (solenoid electromagnets containing cores) installed on the docking panel under different relative distance scales, different actuator types and configurations.

[0005] The present application establishes the following models based on FCNN neural network:

[0006] FCNN-D model: FCNN-D model is used to analyze the spatial distribution of the magnetic field generated by actuators of any type.

[0007] FCNN-F model: FCNN-F model is used to analyze the electromagnetic force / torque of any type of actuator in a known external magnetic field.

[0008] A satellite electromagnetic docking process generalization simulation method based on FCNN network, comprising the following steps:

[0009] Step S1, a large number of different types of actuators are given a large number of random excitation currents for electromagnetic finite element simulation, and the spatial distribution of the generated electromagnetic field is obtained to form an initial magnetic field dataset, which is processed to form a usable magnetic field dataset.

[0010] Step S2, build a new fully connected neural network as FCNN-D model, take the actuator type, spatial point position, excitation current as the input of the network, take the magnetic field strength as the output of the network, set the structure, parameter, activation function, loss function of the neural network.

[0011] Step S3, use the dataset obtained in step S1 to train FCNN-D model, judge whether it meets the requirements through evaluation index, if it meets the requirements, save the training results, if it does not meet the requirements, adjust the model and retrain.

[0012] Step S4, a large number of different types of actuators are given a large number of random excitation currents for electromagnetic finite element simulation, and the spatial distribution of the generated electromagnetic field is obtained to form an initial magnetic field dataset, which is processed to form a usable magnetic field dataset.

[0013] Step S5, build a new fully connected neural network as FCNN-F model, take the actuator type, its position and attitude, excitation current as the input of the network, take the electromagnetic force / torque received by the actuator as the output of the network, set the structure, parameter, activation function, loss function of the neural network.

[0014] Step S6, use the dataset obtained in step S4 to train FCNN-F model, judge whether it meets the requirements through evaluation index, if it meets the requirements, save the training results, if it does not meet the requirements, adjust the model and retrain.

[0015] Step S7, build a dynamic model of square satellite electromagnetic docking, set boundary conditions and dynamic simulation step size.

[0016] Step S8, input the actuator model to be analyzed and its initial spatial position and posture into the FCNN-F model, set the excitation current signal input into the FCNN-F model, import the known external magnetic field, solve the electromagnetic force / torque received by the actuator initially, perform dynamic simulation iteration, update the spatial position and posture of the cube star using the output of the FCNN-F model, calculate the new position and posture of the actuator and re-input the FCNN-F model, other inputs remain unchanged, solve the electromagnetic force / torque received by the actuator at the new position, step the dynamic simulation, repeat the above process until the boundary conditions set in step S7 are reached, and determine whether the docking is successful or not.

[0017] Step S9, visualize the electromagnetic docking motion process of the cube star calculated in step S8.

[0018] In step S1, the actuator model parameters include the installation position of the center positions of the four solenoid electromagnets on the docking panel of the cube star, represented by four two-dimensional coordinates; the inner and outer diameters of the solenoid, the length of the solenoid, the number of turns of the coil, the core material, the core diameter, and the core length can be represented by a 15-dimensional vector. The excitation current includes four actuator excitation currents, which can be represented by a 4-dimensional vector. The spatial electromagnetic field distribution can be represented by a set of mappings of corresponding spatial point three-dimensional coordinates and three-dimensional magnetic field strength vectors at the target cube star.

[0019] A large number of simulations are performed using known finite element simulation methods of electromagnetics, and each time a three-dimensional model is established according to randomly set actuator model parameters and excitation currents, and the corresponding spatial electromagnetic field distribution simulation results are saved.

[0020] After classifying and binding the magnetic field strength and corresponding spatial point position in the initial magnetic field data set, the actuator model and excitation current are randomly arranged and normalized to form a usable magnetic field data set.

[0021] In step S2, the fully connected neural network model adopts a supervised learning method, which is composed of an input layer, first, second, third, and fourth hidden layers, and an output layer.

[0022] The input layer includes 22 input nodes corresponding to the actuator model parameters, excitation current, and spatial point position.

[0023] Each of the four hidden layers includes 512 neurons, and the activation function uses the ReLU function.

[0024] The output layer includes three output nodes corresponding to the three dimensions of the magnetic field strength vector.

[0025] The neural network parameters include learning rate, batch size, training rounds, and dropout rate, and the loss function uses cross-entropy loss.

[0026] In step S3, the training process is to train the neural network with 60% of the data in the available magnetic field data set obtained in step S1, update the weights and biases of each neuron in the four hidden layers, test the model with the remaining 40% data, and use mean square error (MSE) as the evaluation index. When the relative error is less than 5%, the model is considered to meet the accuracy requirements, otherwise the model layer, the number of neurons in each layer and the model parameters are adjusted and retrained to meet the requirements.

[0027] In step S4, the executor space position can be represented by the three-dimensional coordinates of the geometric center of the executor in the global coordinate system, and the executor space attitude can be represented by the Euler angles of the body of the cube star carrying the executor relative to the global coordinate system. The electromagnetic force / torque received by the executor can be represented by a six-dimensional vector containing the three dimensions of the electromagnetic force vector and the three dimensions of the electromagnetic torque vector, and the executor should belong to the tracking cube star.

[0028] A large number of simulations are performed using known electromagnetic finite element simulation methods. Each time, a three-dimensional model is established according to the randomly set executor model parameters, excitation current, executor space position and attitude, and the corresponding electromagnetic force / torque simulation results of the executor are saved.

[0029] The known external magnetic field is the electromagnetic field distribution obtained by matching search in the initial magnetic field data set in step S1 according to the executor model parameters and excitation current set in each electromagnetic finite element simulation analysis. The executor parameters and excitation current of the target cube star and the tracking cube star (hereinafter referred to as the target star and the tracking star) are set to the same value, which is considered to be unified and symmetrical for the cube star cluster.

[0030] After classifying and binding the electromagnetic force / torque received by the executor in the initial force field data set, the corresponding executor space position and attitude, and the set executor model and excitation current, they are randomly arranged and normalized to form a force field data set.

[0031] In step S5, the fully connected neural network model adopts a supervised learning method, which is composed of an input layer, first, second, third and fourth hidden layers, and an output layer.

[0032] The input layer includes 25 input nodes corresponding to the executor model parameters, excitation current, and space position and attitude.

[0033] The four hidden layers each include 512 neurons, and the activation function uses the ReLU function.

[0034] The output layer includes six output nodes corresponding to the electromagnetic force / torque received by the executor.

[0035] The neural network parameters include learning rate, batch size, training round number, dropout rate, and the loss function adopts cross-entropy loss.

[0036] In step S6, the training process is to train the neural network with 60% of the data in the available force field data set obtained in step S4, update the weights and biases of each neuron in the four-layer hidden layer, test the model with the remaining 40% data, and use mean square error (MSE) as the evaluation index. When the relative error is less than 5%, the model is considered to meet the accuracy requirement, otherwise the model layer, the number of neurons in each layer and the model parameters are adjusted and retrained to meet the requirements.

[0037] In step S7, the cubic star electromagnetic docking dynamics model adopts a simple Newtonian motion law model for ground tests facing a fixed target star. The fixed target star is used to facilitate the study of the relative motion of the tracking star. The key parameters of force and torque in Newtonian motion law are given by the FCNN-F model.

[0038] The boundary conditions include setting a simulation time sufficient to complete the entire simulation of the cubic star electromagnetic docking motion, setting an index for considering the docking of the cubic star to be successful and terminating the simulation program in advance, and setting an index for considering other special cases such as scratching and collision of the cubic star to forcibly terminate the simulation program.

[0039] In step S8, the training of FCNN-D and FCNN-F should be completed when this step is performed.

[0040] The excitation current sequence is a time sequence of the excitation current of an actuator of interest. The length of the sequence depends on the set dynamics simulation time and step size. Each simulation step should correspond to an excitation current (four-dimensional vector) in chronological order. The excitation current at the kth step of simulation is denoted as i(k).

[0041] The following process is performed when the electromagnetic docking simulation is performed to the kth step:

[0042] The actuator model to be analyzed is input into the FCNN-D model, and i(k) is input. A set of spatial position samples in a large enough area is generated according to the spatial position data density obtained in step S1. Each spatial position generated is input into the FCNN-D model, and the FCNN-D model is repeatedly run to calculate the spatial distribution of the electromagnetic field generated by the target star actuator. It should be a set of mappings of a series of generated sampling point three-dimensional coordinates and the magnetic field strength three-dimensional vector at that place, denoted as H(k), where the magnetic field strength at a specific position is denoted as H(k, x, y, z).

[0043] The actuator model to be analyzed, the spatial position and attitude of the tracking star are input into the FCNN-F model, i(k) is input, H(k) calculated in process A is input as the known external magnetic field, the spatial position and attitude of the tracking star at the kth step of simulation are recorded as p(k), and the FCNN-F calculation result is recorded as F(k).

[0044] According to the established electromagnetic docking dynamics model of the cubic star, the change of the spatial position and attitude of the tracking star under the action of F(k) is calculated by inertia integration and recorded as δp(k), and the six-degree-of-freedom velocity increment of the tracking star is recorded as δv(k), so that the position and attitude of the tracking star at the k+1th step of simulation can be obtained as p(k+1) = p(k) + δp(k), and the six-degree-of-freedom velocity is v(k+1) = v(k) + δv(k).

[0045] Processes A, B and C are repeated until the boundary condition is reached, and the time series of the spatial position and attitude of the tracking star and the velocity can be obtained.

[0046] In step S9, according to the time series of the spatial position and attitude of the tracking star and the velocity obtained in step S8, and the given cubic star model, the motion of the two cubic stars during the electromagnetic docking process can be visualized, and whether the docking is successful and the docking quality can be judged.

[0047] Compared with the prior art, the advantages of the present application are as follows:

[0048] The satellite electromagnetic docking process generalization simulation method based on the full connection neural network has high generalization ability, and the neural network trained by high-precision finite element simulation data can give high-precision electromagnetic fields generated by the actuator and the interaction between the actuators without re-performing complex finite element analysis when the platform simulates and analyzes new actuators, which facilitates the comparison of electromagnetic performance between different models of actuators and improves the efficiency of actuator design.

[0049] The interaction between neural networks and the joint simulation of neural networks and numerical analysis platforms organically combines the actuator model and the interaction between the tracking star and the electromagnetic actuators on the target star, realizes the unification of the cubic star electromagnetic docking management, and at the same time establishes a high-precision electromagnetic docking dynamics model suitable for the full spatial scale of the actual electromagnet, which further enhances the generalization ability and universality of the method. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 It is a flowchart of a cubic star electromagnetic docking process generalization simulation analysis method;

[0051] Figure 2 It is a double-star docking dynamics coordinate system diagram of a specific embodiment of the present application;

[0052] Figure 3 Figure 1 is a schematic diagram of modeling an electromagnetic actuator based on Ansys Maxwell software according to an embodiment of the present application; DETAILED DESCRIPTION

[0053] The present application will be described in detail below with reference to the accompanying drawings.

[0054] As shown in Figure 1 , the method according to the present application is implemented as follows:

[0055] (1) As shown in Figure 2 , define a global motion reference coordinate system S1, a target star fixed coordinate system S t , and a tracking star fixed coordinate system S c . Define the spatial position and attitude of the actuator or tracking star in S1, and define the yOz plane of the cubic star fixed coordinate system as the installation plane of the actuator, and define the two-dimensional coordinates of the electromagnet installation position in the plane.

[0056] (2) Corresponding to steps S1, S2, and S3, build an FCNN-D model with a fully connected neural network structure, perform electromagnetic finite element simulation to obtain a data set, and train and optimize the FCNN-D model.

[0057] Perform electromagnetic finite element simulation analysis in Ansys Maxwell to obtain the FCNN-D training data set:

[0058] A1, use the known Ansys Maxwell electromagnetic finite element simulation analysis process to model and simulate a certain actuator, wherein 15 dimensions in the actuator model parameters and 4 dimensions of the excitation current are defined in the form of global variables;

[0059] A2, generate at least 50,000 random 15-dimensional vectors within a reasonable range after deduplication, with the vector dimensions corresponding to the 15 dimensions in the actuator model parameters, and the same number of random 4-dimensional vectors after deduplication, with the vector dimensions corresponding to the 4 dimensions in the excitation current. The above two random vector sets will be used as the list set for corresponding global variable parameterization scanning, and the same number of random 3-dimensional vectors after deduplication will be generated, with the vector dimensions corresponding to the 3 dimensions of the spatial point position. Merge the three types of vectors in random order to obtain a set of at least 50,000 random 22-dimensional vectors, which will be used as the data set input part of the FCNN-D model.

[0060] A3, start to perform electromagnetic finite element simulation analysis, traverse the electromagnetic field space distribution generated by the random type actuator under the action of the random excitation current obtained in B to save respectively, the magnetic field intensity vector at any position in the space can be read from the electromagnetic field space distribution simulation result set, and the set of electromagnetic field space distribution simulation results is recorded as a field database.

[0061] A4, traverse all 19-dimensional vectors in the FCNN-D model data set input part, search in the field database to obtain the magnetic field intensity at a random position in the electromagnetic field generated by a random type actuator under the action of a random excitation current, and then a set of at least 50,000 three-dimensional magnetic field intensity vectors can be obtained, which will serve as the data set output part of the FCNN-D model.

[0062] Train the above FCNN-D model and optimize the model according to the training result:

[0063] Randomly select 60% of the data in the FCNN-D model data set as the training set of the FCNN-D model to train the model. The remaining 40% data is used as the validation set of the FCNN-D model, and the mean square error (MES) is used as the evaluation index to test the training effect. The training of the full connection neural network is a known technology, and will not be described in detail in the present application. The training effect of the FCNN-D model is investigated, and when the relative error is less than 5%, the model is considered to meet the accuracy requirement, otherwise the number of hidden layers, the number of neurons in each layer and the model parameters are adjusted and retrained to meet the requirements.

[0064] (3) Corresponding to steps S4, S5 and S6, the FCNN-F model is built with the structure of the full connection neural network, the data set is obtained by electromagnetic finite element simulation, and the FCNN-F model is trained and optimized.

[0065] The electromagnetic finite element simulation analysis is performed in Ansys Maxwell to obtain the data set for training FCNN-F:

[0066] B1, use the known Ansys Maxwell electromagnetic finite element simulation analysis process to model and simulate the actuator, wherein the 15 dimensions of the actuator type parameter, the 4 dimensions of the excitation current, the 3 dimensions of the center point space coordinates of the actuator mounting panel, and the 3 dimensions of the Euler angle of the actuator mounting plane normal vector relative to the unit vector of the global coordinate system x axis are defined in the form of global variables by parameterized scanning. The specific position of each actuator can be solved by simple constraints and known attitude calculation formula during modeling.

[0067] As Figure 3As shown, the tracking star executor spatial position and attitude are established, and the electromagnetic executor at the lower part of the figure is only used for position reference on the target star executor and does not apply excitation current to it.

[0068] B2, generate at least 50,000 random 15-dimensional vectors after deduplication, and the vector dimensions correspond to the 15 dimensions in the executor model parameters respectively; the same number of random 4-dimensional vectors after deduplication, and the vector dimensions correspond to the 4 dimensions in the excitation current respectively; and the same number of random 6-dimensional vectors after deduplication, and the vector dimensions correspond to the 6 dimensions of the executor spatial position and attitude. The above three random vector sets will be used as the list set of the corresponding global variable parameterization scanning. The three types of vectors are merged in random order to obtain a set formed by at least 50,000 random 25-dimensional vectors, which will be used as the data set input part of the FCNN-F model.

[0069] B3, use a known process to write an Ansys Maxwell finite element simulation recording script, add a part to read an external magnetic field file in the script, and search for the spatial electromagnetic field distribution according to the executor model and excitation current traversed in the B2 vector set in the field database obtained when establishing the FCNN-F model, and import the file into the simulation scene as the known external magnetic field.

[0070] B4, start electromagnetic finite element simulation analysis, save the electromagnetic force / torque respectively received by a random model executor under the action of a random excitation current at a random spatial position and attitude in the known electromagnetic field generated by the same model executor under the action of the same excitation current, and then obtain a set formed by at least 50,000 6-dimensional electromagnetic force / torque vectors, which will be used as the data set output part of the FCNN-F model.

[0071] Train the FCNN-F model and optimize the model according to the training results.

[0072] (4) Corresponding to step S7, set the simulation parameters and perform dynamic modeling.

[0073] Set the simulation parameters, i.e., the boundary conditions, set the simulation time to 10s, set the simulation step to 0.01s, and set the condition for successful docking when the four pairs of electromagnets in the two cubic star electromagnetic docking executors are in contact with each other, i.e., the distance is close to 0 and the radial misalignment of each pair of electromagnets is not more than 2mm. Detect the envelopes of the two cubic stars, and when the envelopes intersect but do not reach the successful docking condition, consider that an accident has occurred, at which time the simulation will be forcibly terminated. When it is detected that the simulation reaches the limit time or the cubic star motion reaches a far enough distance to cause the interaction between the executors to be close to 0, the simulation is terminated.

[0074] Corresponding to step S8, based on the FCNN-D model and the FCNN-F model optimized by training, a numerical analysis platform such as Python is used to jointly simulate the electromagnetic docking process of the CubeSat to be analyzed, and the motion of the whole process is obtained by step iteration, and whether the docking is successful or not is judged.

[0075] The specific process of the joint simulation is as follows:

[0076] The time sequence of the excitation current to be analyzed is generated, and the excitation current of the kth simulation step is denoted as i(k). In the examples of the present application, a step current signal with an amplitude of 1A is selected for all electromagnets, and then the excitation current in each simulation step is i(k) = [1 1 1 1] T (unit: A), and then the excitation current time sequence can be expressed as

[0077] The joint simulation is started, and when the simulation is stepped once (when the simulation is stepped to the kth step), the following operations are performed:

[0078] C1, input the actuator model to be analyzed into the FCNN-D model, and input i(k), generate a sufficient number of sampling point spatial position sets in a large area according to the sampling point density in the electromagnetic field spatial distribution file obtained when the FCNN-D model dataset is obtained, input each generated spatial position into the FCNN-D model, repeatedly run the FCNN-D model, calculate the electromagnetic field spatial distribution generated by the target star actuator, save it as a csv file, which is a series of generated sampling point three-dimensional coordinates and the mapping set of the three-dimensional vector of the magnetic field strength at this point, denoted as H(k), representing the spatial magnetic field distribution generated by the target star under the action of the actuator excitation current at the kth step.

[0079] C2, input the actuator model to be analyzed, the tracking star spatial position and attitude into the FCNN-F model, and input i(k), import H(k) as the known external magnetic field, the tracking star spatial position and attitude of the kth simulation step are denoted as p(k), and the FCNN-F calculation result is denoted as F(k).

[0080] C3, according to the established cubic satellite electromagnetic docking dynamics model, the spatial position and attitude increment of the tracking satellite under the action of F(k) is calculated by inertia integration and recorded as δp(k), the six-degree-of-freedom velocity increment of the tracking satellite is recorded as δv(k), then the position and attitude of the simulation tracking satellite at the k+1 step can be obtained as p(k+1) = p(k) + δp(k), and the six-degree-of-freedom velocity of the simulation tracking satellite at the k+1 step can be obtained as v(k+1) = v(k) + δv(k), so that the alternating step closed-loop joint simulation process between the interactive neural network model and the data analysis platform is realized. In the examples of the present application, python is used for numerical analysis of dynamics. The spatial position and attitude of the cubic satellite at the first simulation step and the six-degree-of-freedom velocity are simulation initial conditions, which are artificially set as [200 cm 20 cm-30 cm 10°-20° 15°],

[000] m / s,

[000] deg / s.

[0081] C4, repeat the above operation until the simulation boundary condition is reached, and then the spatial position and attitude of the cubic satellite and the six-degree-of-freedom velocity at each simulation step have been calculated, and then the six-degree-of-freedom motion time series of the tracking satellite can be obtained as

[0082] Detect the simulation termination condition triggered in the above simulation process, and judge whether the docking is successful.

[0083] (5) Corresponding to step S9, the motion full process obtained in (4) is visualized.

[0084] The visualization process is to read the six-degree-of-freedom time series of the tracking satellite obtained by joint simulation, and use the built-in drawing software of python to draw the spatial position and attitude change waveform and the six-degree-of-freedom velocity waveform of the tracking satellite.

Claims

1. A generalized simulation method for satellite electromagnetic docking process based on FCNN, characterized in that: The method includes the following steps: Step S1: Apply a large number of random excitation currents to a large number of actuators of different models and perform electromagnetic finite element simulation to obtain the spatial distribution of the electromagnetic field generated to form an initial magnetic field dataset. After processing, a usable magnetic field dataset is formed. Step S2: Build a new fully connected neural network as the FCNN-D model. Use the actuator model, spatial point location, and excitation current as the network input, and the magnetic field strength as the network output. Set the structure, parameters, activation function, and loss function of the neural network. Step S3: Use the dataset obtained in step S1 to train the FCNN-D model and use the evaluation metrics to determine whether the requirements are met. If the requirements are met, save the training results; otherwise, adjust the model and retrain. Step S4: Perform electromagnetic finite element simulation on a large number of actuators with different models and excitation currents to obtain the electromagnetic force received by them at different positions and attitudes in the external magnetic field and the electromagnetic force rectangle to form an initial force field dataset. After processing, a usable force field dataset is obtained. Step S5: Build a new fully connected neural network as the FCNN-F model. Take the actuator model, its position and orientation, and the excitation current as the network inputs, and take the electromagnetic force and electromagnetic torque on the actuator as the network outputs. Set the structure, parameters, activation function, and loss function of the neural network. Step S6: Train the FCNN-F model using the dataset obtained in step S4. Determine whether the requirements are met by the evaluation metrics. If the requirements are met, save the training results. If not, adjust the model and retrain. Step S7: Build a dynamic model of the electromagnetic docking of the satellite and set boundary conditions and dynamic simulation step size; Step S8: Input the actuator model to be analyzed and its initial spatial position and attitude into the FCNN-F model, set the excitation current signal to be input into the FCNN-F model, import the known external magnetic field, solve the electromagnetic force and electromagnetic torque initially experienced by the actuator, perform dynamic simulation iteration, update the CubeSat spatial position and attitude using the output of the FCNN-F model, calculate the new position and attitude of the actuator and re-input it into the FCNN-F model, keep other inputs unchanged, solve the electromagnetic force and electromagnetic torque experienced by the actuator at the new position, perform step dynamic simulation, repeat the above process until the boundary conditions set in step S7 are reached, and determine whether the docking is successful. Step S9: Visualize the CubeSat electromagnetic docking motion process calculated in step S8.

2. The method of claim 1, wherein the method is based on a FCNN. The FCNN-D model and the FCNN-F model mentioned in steps S2 and S5 have data interaction rather than exist independently. Specifically, the external magnetic field data on which the FCNN-F model relies for calculation needs to be calculated and provided by the FCNN-D model.

3. The method of claim 1, wherein the method further comprises: A full connection neural network model is constructed to learn the complex nonlinear mapping relationship in the electromagnetic field and directly output the spatial magnetic field distribution of the actuator, as well as the electromagnetic force and torque suffered by the actuator in the external magnetic field. The data used for training the full connection neural network model is obtained by finite element simulation of electromagnetic analysis on a large number of different types of actuators with a large number of random excitation currents, covering various actuator types and docking working conditions.

4. The method of claim 1, wherein the method is based on a FCNN. The introduced full connection neural network has high generalization ability for models within a reasonable range due to its data-driven characteristics, and can directly give the electromagnetic performance of electromagnetic docking actuators that have not appeared before without complex modeling and analysis processes.

5. The method of claim 1, wherein the method is based on a FCNN. The method adopts a joint simulation framework, establishes a rigid body kinematics simulation model in the numerical analysis platform, and establishes an electromagnetic calculation model under the neural network framework, which is formed by FCNN-D neural network and FCNN-F neural network in series. FCNN-D neural network is used to calculate the magnetic field intensity, and then FCNN-F neural network is connected in series. FCNN-F neural network calculates the electromagnetic force and electromagnetic torque suffered by the actuator according to the calculation result of FCNN-D neural network. The above rigid body kinematics simulation model and electromagnetic calculation model interact with each other, including the numerical analysis platform transmitting the pose information of the simulated cubic star and electromagnetic docking actuator to the electromagnetic calculation model, and the electromagnetic calculation model transmitting the interaction between electromagnets to the numerical analysis platform. In this process, the pose output by the last step simulation is used as the pose information of the electromagnetic docking actuator in the electromagnetic simulation of the electromagnetic calculation model in the current step simulation, and the calculation result of the electromagnetic calculation model is used as the external force and torque of the dynamic model in the current step simulation to update the pose information of the cubic star in the current step simulation, and then the signal flow forms a dynamic closed loop structure of step iteration.

6. The method of claim 1, wherein the method is based on a FCNN. In the joint simulation process, the excitation current is input into the electromagnetic calculation model in the form of time sequence, and then the external control system is allowed to control the current according to the preset control law to change the electromagnetic current value of each step simulation. The control law is the control method used for specific working condition simulation.

Images