Enhancing space sensor resilience with transfer learning in data-scarce scenarios

Abstract In Mars exploration missions, harsh environmental conditions, such as those generated by dust devils, can damage sensing systems. Soft sensors offer a promising solution in such scenarios, but their implementation is challenging when data is scarce. This paper explores the use of transfer learning (TL) to enhance sensor resilience, specifically addressing the issue of limited data availability. First, pre-trained models are developed using wind sensor data from the Temperature and Wind Sensors for the TWINS instrument (NASA InSight Mission), serving as the source domain for this study. These models account for the malfunction of a single wind sensing board or transducer. The pre-trained models are then adapted using TL to the Mars Environmental Dynamics Analyzer (MEDA) wind sensor (NASA Perseverance Mission), which shares common sensing principles, and which suffered a malfunction during the mission. Hyperparameter tuning further improves the performance of the TL models, yielding better results than models trained solely on a small MEDA dataset. The results demonstrate the effectiveness of the TL-based approach in recovering variables from the MEDA wind sensor despite partial failures and data limitations. Overall, the TL-based method improves performance by 10.21%–22.04% compared to models trained exclusively on the limited MEDA dataset.