Fault Diagnosis with Spacecraft High-Dimensional Data Based on Machine Learning

In order to ensure the safe, reliable and stable operation of spacecraft, this paper studies the fault diagnosis of spacecraft high-dimensional data by applying the machine learning technology, aiming at the characteristics of spacecraft, such as closed-loop fault propagation, few fault samples and limitation of on-orbit computing capability. For the experimental data under the typical failure modes of spacecraft sensors and actuators, this paper will use some machine learning algorithms like K-Nearest Neighbor (KNN), Bayesian classifier, KNN based on Principal Component Analysis (PCA+KNN), to analyze fault diagnosis and to make a comparative analysis of the results. The results show as follows: it can reach better accuracy by using supervised learning algorithms in fault diagnosis classification; the dimensionality reduction can accelerate the calculation, but at the same time it will lose diagnostic accuracy, especially for highly-coupled and high-dimensional data; in the algorithm design, the task requirements and application conditions should be considered comprehensively, meanwhile the optimal fault diagnosis model can be selected by balancing the accuracy and computing time.