Intelligent Identification Of Inertia Matrix Of Combined Spacecraft In On-orbit Capturing

With the development of space missions,there is an increasing number of out-of-control spacecraft due to fuel exhaustion or system failure,the on-orbit service technologies for out-ofcontrol spacecraft have attracted extensive attention from domestic and foreign researchers.In the process of capturing the target,the inertia matrix of a combined spacecraft composed of space robot and target changes greatly compared with that of the space robot before the capture begins,which may lead to the inability of the space robot to effectively control the attitude of the whole combined spacecraft.Therefore,it is very important to identify the inertia matrix of combined spacecraft accurately.The inertia matrix identification of combined spacecraft can be divided into two cases,one is to identify the inertia matrix of combined spacecraft with a fixed configuration,which is a problem of time-invariant system parameter identification.The other is to identify the changing inertia matrix of combined spacecraft from the perspective of the whole capture process,which is a problem of time-varying system parameter identification.In this paper,the above two identification problems of the inertia matrix of combined spacecraft are deeply analyzed and studied.In addition,it will face some difficulties in the actual execution of the identification task,for example,the inertia matrix identification accuracy of combined spacecraft depends on the accuracy of the measured angular rate,and the measurement noise of the angular rate measured value may be very complex in some disturbance conditions;for the real-time identification task,the inertia matrix changes smoothly with time and also occurs unpredictable mutation at the moment when the target is captured.All these cases further increase the difficulties of accurately identifying the inertia matrix of combined spacecraft.Because of this,the deep learning methods in artificial intelligence technology are applied to the researches of inertia matrix identification in this paper.From the perspectives of the timeinvariant system and time-varying system,the identification methods of inertia matrix are studied when complex measurement noise exists in the angular rate measurement data of the combined spacecraft,and the inertia matrix changes with time and changes abruptly in the process of capturing the target.The main contents are given as follows:(1)The on-orbit identification method of the time-invariant inertia matrix of combined spacecraft is studied.To improve the identification performance of the inertia matrix of combined spacecraft with fixed configuration in the environment of complex measurement noise,an identification method of the time-invariant inertia matrix based on deep learning is proposed.Firstly,the attitude dynamic model of a fixed configuration combined spacecraft consisting of a space robot and a captured target is established.According to the dynamic model,the domain randomization method is used to generate the angular rate,control moment,and the corresponding inertia matrix of the combined spacecraft,which is used to simulate the actual measurement data for constructing the training set and testing set.Then the deep neural network(DNN)is constructed and trained by a training set and a designed training strategy.Finally,the identification effectiveness of the DNN model is tested using two testing sets with different ranks of measurement noises,and compared with the recursive least squares(RLS)identification method and the traditional DNN trained by traditional training strategy.(2)An on-orbit identification method of the time-invariant inertia matrix of combined spacecraft based on ensemble deep learning is investigated.Because a single DNN will fall into different local optimal solutions when trained by the gradient descent method,the performance of the trained DNN for parameter identification is not stable.To solve the above problems,an ensemble deep learning method based on least squares is proposed to realize the high-accuracy identification of the inertia matrix of combined spacecraft in the environment of complex measurement noise.In this method,a single DNN is constructed as an individual learner of the ensemble deep learning method.After obtaining a certain number of individual learners by using the injecting randomness method,the output of all individual learners is with six linear regression functions,and the least square method is used to automatically determine the optimal regression coefficients of these linear regression functions.Finally,two testing sets with different ranks of measurement noises are used to test the identification performance of the proposed method and compared with the existing RLS identification method,single DNN model,and averaging ensemble deep learning,etc.Simulation results show that the proposed ensemble deep learning method based on least squares has obvious advantages in identification accuracy and stability.(3)The on-orbit identification method of the time-varying inertia matrix of combined spacecraft in the process of the space robot capturing a target is studied.Aiming at the problem that the inertia matrix of combined spacecraft changes with time during space robot capturing the target and occurs mutation at the moment when the target is captured,a real-time identification method of inertia matrix based on long short-term memory(LSTM)is proposed.Firstly,the dynamic models of the space robot and combined spacecraft are developed according to the two stages of pre-capture and post-capture.Based on the dynamic models and the domain randomization method,sufficient data is then obtained for training and testing.The deep parameter identification network,composed of an LSTM network and multi-layer fully connected network,is trained using a training set and a designed training strategy.Finally,a group of testing data is used to test the identification effectiveness of the trained network.Simulation results show that this method can accurately identify the time-varying inertia matrix of the combined spacecraft.(4)Based on the axisymmetric case of rigid body attitude dynamics,the operation mechanism of DNN used for identifying parameters is analyzed,and a parameter generation method for rapidly constructing DNN is proposed.The operation mechanism of DNN in the parameter identification task is analyzed from the perspective of the internal hidden layer.In the analysis of the hidden layers of DNN,taking the rigid body attitude dynamics as an example,some statistical relations between the parameters of the hidden layers of DNN and their input are revealed.These statistical relations are further verified in system parameter identification of damped harmonic oscillator and damped pendulum.By imitating these statistical relations,a parameter generation method for rapidly constructing DNN is proposed,and the validity of the method is verified using the testing sets.