| With the continuous development of the micro-nano satellites control technology,the attitude measurement and control system is getting more and more complex,and the coupling between system parameters is getting enhanced.Traditional fault diagnosis methods are increasingly inadequate in the face of complex control systems,and the more intelligent fault diagnosis methods are urgently needed to improve the reliability of attitude measurement system of micro-nano satellites.Machine learning has great advantages in dealing with complex systems and strong coupling parameters,so this paper introduces machine learning into fault diagnosis field of micro-nano satellites.This paper mainly does the following work:Firstly,the shortcomings of the existing technology are analyzed.According to the on-orbit fault data of micro-nano satellites,the fault modes of attitude sensors are summarized.Further more,the measurement models and the fault models of the attitude sensors are established.Secondly,aiming at the shortcomings of the existing fault diagnosis technology,which are weak in extracting features of attitude sensors measurements and dealing with unknown faults,a fault diagnosis algorithm based on gated-recurrent unit is proposed.This paper adds peephole connection and bidirectional stack structure to make better use of the historical and future data,then outputs the hidden features of attitude sensors measurements to analyze the unknown faults.Simulation experiments show that this algorithm has great predictability in attitude sensors measurements,as well as great identifiability in various faults of attitude sensors.This algorithm improved the ability of existing technology to handle unknown faults.Thirdly,The simulation results of the designed fault diagnosis algorithm based on the gated-recurrent unit show that it has the problem of insufficient generalization and consumes too much on-orbit computing resources when dealing with known faults.Referring to Autoencoder structure,the attitude sensor measurements are deconstructed and reconstructed to improve the generalization of fault diagnosis algorithm.Meanwhile,the softmax function is added to directly classify known faults to save computing resources.Simulation results show that the improved algorithm improves the precision by 3.17%,consumes less computing resources and has better ability to deal with faults.Fourthly,to verify the engineering validity of the two proposed algorithms,the designed fault diagnosis algorithms are transplanted to the semi-physical simulation platform to verify the predictability.The experiments show that the two algorithms still have good predictive performance in the presence of unknown environmental interference,and the recall of the two algorithms can still up to 90%.In this paper,machine learning is introduced to improve the ability of micro-nano satellites to extract features of attitude sensors measurements.It makes up for the shortages of existing technology to deal with unknown faults and provides a new technical approach to improve the reliability of attitude measurement system of micro-nano satellites. |
