Battery State of Health Estimation by Means of Latent Space Tracking: Applications to Satellite Subsystems

In modern space missions, ensuring the reliability of the onboard Electrical Power Subsystem (EPS) is critical for mission safety, continuity, and longevity. Lithium-ion (Li-ion) batteries, thanks to their high energy density and capacity-to-weight ratio, are key components of the EPS. However, their performance declines over time due to electrochemical aging, leading to reduced capacity and higher failure risk, especially problematic in space, where data are limited, full discharge cycles are rare, and physical intervention is impossible after launch. This thesis proposes a data-driven method to estimate the State of Health (SoH) of Li-ion batteries by leveraging latent representations learned via autoencoders. The developed pipeline uses Long Short-Term Memory (LSTM) networks to compress discharge cycles into latent vectors, which are then used for both SoH regression and latent space tracking. Regularization techniques are applied to improve model generalization, while subspace tracking algorithms, GROUSE and PAST, model the dynamic evolution of the latent space. This method was validated using the NASA Battery Data Set (2008), available through the Prognostics Data Repository. Training was performed on discharge cycles from three batteries (B0005, B0006, B0007), while testing was carried out on a fourth, previously unseen battery (B0018), using both full and partial discharge cycles. The results show that the method can effectively monitor battery degradation over time and predict future capacity with strong robustness, even under partial observability and noisy conditions. Moreover, latent space analysis makes it possible to capture not only degradation trends but also partial regeneration phenomena, often observed in real systems but difficult to model using traditional techniques. This approach offers a promising tool to support decision-making, both onboard and on the ground, enabling effective monitoring of satellite health and end-of-life planning.