Mission control and fault management are fundamental in safety-crit- ical scenarios such as space applications. To this extent, fault detection tech- niques are crucial to meet the desired safety and integrity level. This work pro- poses a fault detection system exploiting an autoregressive model, which is based on a Deep Neural Network (DNN). We trained the aforementioned model on a dataset composed of telemetries acquired from Mars Advanced Radar for Sub- surface and Ionosphere Sounding (MARSIS). The training process has been de- signed as a sequence-to-sequence task, varying the length of input and output time series. Several DNN architectures were proposed, using both Long Short- Term Memory (LSTM) and Gated Recurrent Unit (GRU) as basic building blocks. Lastly, we performed fault injection modeling faults of different nature. The results obtained show that the proposed solution detects up to 90% of in- jected faults. We found that GRU-based models outperform LSTM-based ones in this task. Furthermore, we demonstrated that we can predict signal evolution without any knowledge of the underlying physics of the system, substituting a DNN to the traditional differential equations, reducing expertise and time-to-market concern- ing existing solutions.
Fault Detection Exploiting Artificial Intelligence in Satellite Systems
Ferrante, Nicola;Giuffrida, Gianluca;Nannipieri, Pietro;Bechini, Alessio;Fanucci, Luca
2023-01-01
Abstract
Mission control and fault management are fundamental in safety-crit- ical scenarios such as space applications. To this extent, fault detection tech- niques are crucial to meet the desired safety and integrity level. This work pro- poses a fault detection system exploiting an autoregressive model, which is based on a Deep Neural Network (DNN). We trained the aforementioned model on a dataset composed of telemetries acquired from Mars Advanced Radar for Sub- surface and Ionosphere Sounding (MARSIS). The training process has been de- signed as a sequence-to-sequence task, varying the length of input and output time series. Several DNN architectures were proposed, using both Long Short- Term Memory (LSTM) and Gated Recurrent Unit (GRU) as basic building blocks. Lastly, we performed fault injection modeling faults of different nature. The results obtained show that the proposed solution detects up to 90% of in- jected faults. We found that GRU-based models outperform LSTM-based ones in this task. Furthermore, we demonstrated that we can predict signal evolution without any knowledge of the underlying physics of the system, substituting a DNN to the traditional differential equations, reducing expertise and time-to-market concern- ing existing solutions.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
