Abstract
Accurate State of Health (SOH) estimation is critical for the reliable operation of satellites during solar eclipse periods. Traditional point-prediction methods often fail to account for the inherent uncertainties and the "capacity recovery effect" observed in lithium-ion batteries. This paper proposes an interval prediction method based on Sample Entropy (SampEn) and a Lower Upper Bound Estimation (LUBE) neural network. By utilizing SampEn to capture subtle capacity fluctuations and optimizing the network with a simulated annealing algorithm, the proposed approach provides robust prediction intervals that prioritize mission safety.
Introduction
For satellites, the battery system is the sole power source during "shadow regions" when solar panels are inactive. As lithium-ion batteries age, their capacity degrades non-linearly, often exhibiting temporary increases known as the capacity recovery effect. While most models provide single-point estimates of SOH, these are often brittle in the face of sensor noise and environmental variability. For mission operators, understanding the uncertainty—specifically the lower bound of guaranteed power—is more vital than a simple average estimate.
Methodology
The proposed framework consists of a multi-stage pipeline designed to produce sensitive and reliable prediction intervals.
Feature Extraction
The model utilizes two primary features: discharge cutoff voltage and Sample Entropy (SampEn). SampEn serves as a measure of signal complexity and regularity. Unlike standard voltage metrics, SampEn is highly sensitive to local fluctuations, allowing it to effectively characterize the capacity recovery effect where lithium ions are regenerated during rest periods.
The LUBE Model and Optimization
The core of the estimation is a Lower Upper Bound Estimation (LUBE) neural network. Unlike standard architectures, LUBE features two output neurons that directly predict the upper and lower boundaries of the SOH interval.
To train this network, a modified Coverage Width-based Criterion (CWC) is used as the loss function. This criterion balances the Prediction Interval Coverage Probability (PICP)—the frequency with which the true value falls within the interval—against the Normalized Mean Prediction Interval Width (NMPIW). Because the CWC is non-differentiable, a Simulated Annealing algorithm is employed to optimize the network weights.
Experimental Setup and Data
The study validates the method using the NASA Ames Prognostics Center of Excellence (PCoE) lithium-ion battery dataset (specifically battery #18). This dataset includes comprehensive aging data recorded at 24°C.
The analysis focuses on extracting features from the discharge phase to track the degradation process accurately across hundreds of cycles.
Results and Discussion
The optimized LUBE model was compared against a baseline Support Vector Machine (SVM) interval prediction method. Key findings include:
- High Reliability: The model achieved a PICP of approximately 0.84, ensuring that the true SOH remains within the predicted bounds for the vast majority of the battery's life.
- Precision: The NMPIW was reduced to roughly 0.08, providing a narrow, actionable window for operators rather than a broad, vague range.
- Dynamic Tracking: Qualitatively, the interval successfully tracked local "spikes" in capacity caused by recovery effects, which are typically smoothed out or ignored by point-prediction models.
Conclusion
By shifting the focus from point estimates to uncertainty intervals, this research provides a more "honest" assessment of satellite battery health. The integration of Sample Entropy allows the model to perceive chemical regenerations that traditional metrics miss. Future work will address the challenges of limited on-orbit telemetry by developing ground-to-space mapping techniques to enhance real-time deployment.
Summary updated