Sensor reduction of batch-produced satellites based on prediction of machine learning model

Abstract

The extensive use of sensors conflicts with the tight schedules and rapid data-processing needs of batch-produced satellites. Existing sensor reduction methods cannot fit the engineering realities of batch-produced satellites, where some locations are unsuitable for sensors, data reduction overlooks key inter-satellite differences, and precise sensor positioning is difficult. In this paper, a combined method of K-means clustering for sensor reduction and fully neural network (FCNN) for response prediction was proposed. This approach constructs candidate sensors and adjusts data structures to suit batch-satellite engineering realities. Additionally, the effects of model algorithms and training-set data were studied. The main findings are as follows: (1) The proposed combined method of K – means clustering and FCNN is effective and robust for sensor reduction of batch-produced satellites. (2) The numbers of hidden layers and neurons in FCNN as well as the number of retained sensors in K – means clustering have little effect on the results. (3) The training set is preferred to encompass data of every vibration direction and amplitude to ensure robust generalization capacities.