Use of A Data-Driven Approach for Time Series Prediction in Fault Prognosis of Satellite Reaction Wheel

Satellites are complicated systems, and there are many interconnected devices inside a satellite that needs to be healthy to ensure the proper functionality of a satellite. Uncertainty and mechanical failure in different integral parts of the satellite pose the major threat for the satellite to remain fully functional for its expected life span. One of the most common reasons for satellite failure is the reaction wheel (RW) failure. Satellite RW fault prognosis can be formed as a two-step process. In this paper, we study the first step where a data-driven approach is used for forecasting the RW parameters that can be used to predict the remaining useful life (RUL) of a reaction wheel onboard satellite. Autoregressive integrated moving average model (ARIMA) and a type of recurrent neural network (RNN) known as the long short-term memory (LSTM) are used for time-series forecasting in this paper. Both models can predict up to a degree of accuracy, even when limited historical data is available. ARIMA works very efficiently as it can capture a suite of different standard temporal structures in time series. Still, when it comes to accuracy, LSTM provides a better regression outcome for our dataset. The results obtained by the models are very optimistic when model parameters are tuned.