Lightweight on-orbit state-of-health estimation for satellite batteries via hardware constraint-driven knowledge distillation

Accurate state of health estimation for lithium-ion batteries is imperative for the reliability of satellite power systems. However, the deployment of high-precision deep learning models on spacecraft is severely hindered by limited on-board computational resources and the unavailability of complete charging data during orbital operations. To address these challenges, this article proposes a lightweight on-orbit distillation framework tailored for resource-constrained edge devices. First, a two-dimensional aging feature map construction method is introduced, which transforms fragmented time-domain data from partial charging segments into standardized image features, effectively mitigating the dependency on full charge-discharge cycles. Subsequently, a resource-aware knowledge distillation strategy is established. A pre-trained residual network is utilized as the teacher model to transfer robust feature extraction capabilities to a lightweight student model. This student model is constructed via a multi-objective optimization process involving structured pruning and weight factorization, adhering to the maximum available memory and allowable processing time of the target hardware. Experimental results demonstrate that the proposed framework significantly reduces the model parameter count to under one million, representing a 55.4% reduction compared to the lightweight baseline, while maintaining high estimation accuracy with a mean absolute percentage error of 1.08%. This approach presents a viable solution for real-time, high-precision health monitoring in aerospace battery management systems.