Aging States Estimation and Monitoring Strategies of Li-Ion Batteries Using Incremental Capacity Analysis and Gaussian Process Regression

1. Problem Statement

The paper addresses the challenge of estimating the State of Health (SoH) and Remaining Useful Life (RUL) of Lithium-ion (Li-ion) batteries in real-world scenarios where continuous monitoring is unavailable.

• Data Scarcity: Most existing methods rely on extensive, continuous cycling histories. However, practical applications (e.g., second-life batteries, fleet management) often provide only sparse information, such as a single diagnostic charge cycle.
• Cell Variability: There is significant cell-to-cell variance in aging trajectories. A model trained on one cell often fails to generalize to a "previously unseen" cell without extensive retraining.
• Uncertainty Quantification: Standard regression methods often provide point estimates without quantifying the confidence or uncertainty of the prediction, which is critical for safety-critical decisions.

2. Methodology

2.1. Data Source

The study utilizes the Oxford Battery Degradation Dataset, comprising 8 small Li-ion pouch cells (LiCoO2/Graphite) cycled at 40°C until End of Life (EOL), defined as 80% SoH. The dataset includes diagnostic cycles recorded every 100 cycles.

2.2. Feature Extraction: Incremental Capacity Analysis (ICA)

Instead of using raw voltage/current time-series data, the authors employ Incremental Capacity Analysis (ICA) to extract high-information features.

• Transformation: ICA converts voltage plateaus in the charging curve into distinct peaks and valleys in the $dQ/dV$ curve.
• Feature Selection: Five specific features ($F_1$ to $F_5$) were manually selected based on visual inspection and Spearman rank correlation with SoH/RUL:
  • $F_1$: Maximum peak of the IC curve ($IC_{peak}$).
  • $F_2$: Voltage corresponding to $IC_{peak}$.
  • $F_3$: Maximum IC in a specific voltage interval (3.5V–3.65V).
  • $F_4$: Maximum positive slope in that interval.
  • $F_5$: Voltage corresponding to the maximum slope.
• Correlation: $F_1$ ($IC_{peak}$) showed the highest correlation with SoH (0.997) and RUL (0.970).

2.3. Regression Models

The authors compare several machine learning approaches:

• Standard Models: Polynomial regression, Feedforward Neural Networks (FFNN), Support Vector Regression (SVR), and a standard Gaussian Process Regression (GPR) trained on pooled data.
• Proposed Model (GPRn): A Multi-Model Gaussian Process Regression Ensemble.
  • Instead of training one GPR on all data, a separate "expert" GPR is trained for each of the $T$ training cells.
  • For a new test cell, the ensemble combines the predictions of these experts using a weighted mixture of Gaussians.
  • Uncertainty Decomposition: The ensemble variance is decomposed into:
    • Epistemic Uncertainty: Uncertainty due to limited data (average variance of experts).
    • Aleatoric Uncertainty: Uncertainty due to inherent cell-to-cell variability (variance of the experts' means).

2.4. RUL Monitoring Strategy

A closed-loop monitoring strategy is derived to operate a battery from Begin of Life (BOL) to EOL using sparse measurements:

1. Perform a single diagnostic charge cycle.
2. Extract ICA features.
3. Estimate SoH (using SVR) and RUL (using GPRn).
4. Calculate a Conservative RUL ($RUL_{cons}$) by subtracting a safety margin proportional to the predicted uncertainty ($k \cdot \sigma$).
5. Operate the battery for $RUL_{cons}$ cycles before the next measurement.
6. Repeat until $RUL_{cons}$ falls below a minimum threshold or SoH < 80%.

3. Key Contributions

1. GPRn Ensemble: Development of a cell-specific GPR ensemble that explicitly models cell-to-cell variability, outperforming pooled GPR models which tend to overfit or fail to capture heterogeneity.
2. Sparse Data Efficiency: Demonstration that accurate SoH and RUL estimation is possible using only one diagnostic cycle from a previously unseen cell, leveraging data from similar reference cells.
3. Uncertainty-Aware Monitoring: Derivation of a practical monitoring strategy that uses the uncertainty quantification from GPRn to dynamically adjust the interval between diagnostic checks, balancing safety and utilization.
4. Quantitative Benchmarking: A comprehensive comparison of ICA-based features against various ML regressors (SVR, FFNN, GPR) for both SoH and RUL tasks.

4. Results

4.1. Estimation Performance

• SoH Estimation: The mapping from ICA features to SoH is highly consistent across cells.
  • Best Model: SVR achieved the lowest error.
  • Accuracy: Normalized Mean Absolute Error (NMAE) of 1.3%.
  • Note: The GPRn ensemble performed slightly worse than pooled GPR for SoH because SoH mapping is largely cell-invariant; a pooled model benefits from more data points.
• RUL Estimation: The mapping is highly sensitive to cell variability.
  • Best Model: GPRn significantly outperformed all other methods.
  • Accuracy: NMAE of 5.3%.
  • Comparison: Standard pooled GPR suffered from severe overfitting (low training error, high test error), while GPRn successfully generalized by averaging individual cell behaviors.

4.2. Monitoring Strategy Performance

Using the GPRn-based strategy with a tuning parameter $k=2$ (balancing safety and utilization):

• Utilization: The strategy successfully utilized >90% of the available cycle budget for most cells.
• Measurement Burden: On average, only 3 to 5 diagnostic measurements were required over a lifetime of 3,300–5,000 cycles.
• Safety:
  • Overcycling Rate: For $k=2$, the overcycling rate (falling below 80% SoH) was 0 for 7 out of 8 cells.
  • SoH Deviation: The residual SoH error at EOL was kept within ±1% for all cells.
  • Fastest Aging Cell: Even for the fastest aging cell (Cell 4), the strategy resulted in only a minor overcycling of -116 cycles, preventing catastrophic failure.

5. Significance and Conclusion

• Practical Applicability: The study proves that complex battery health monitoring does not require continuous data streams. It is feasible to manage battery fleets with sparse, intermittent diagnostics, which is crucial for second-life applications and legacy systems.
• Safety vs. Efficiency: The proposed strategy provides a mathematical framework to tune the trade-off between maximizing battery usage and ensuring safety via uncertainty quantification.
• Future Directions: The authors suggest that with larger datasets, the ensemble weights ($\pi_i$) could be dynamically adapted based on the specific degradation trend of the test cell, potentially reducing aleatoric uncertainty further. Additionally, the method needs validation under varying operating conditions (temperature, C-rates).

In summary, this paper presents a robust, data-efficient framework for battery health management that leverages the physical interpretability of ICA features and the probabilistic strengths of Gaussian Process ensembles to solve the critical problem of estimating battery life with minimal data.