Accurate estimation of the State of Health (SOH) for lithium-ion batteries is a cornerstone of reliable and safe Battery Management Systems (BMS) in electric vehicles. SOH, commonly defined as the ratio of current maximum capacity to initial nominal capacity, directly reflects the degree of performance degradation and aging. As a researcher dedicated to advancing BMS technologies, I have focused on developing precise, non-invasive, and computationally efficient SOH estimation methods. This work presents an integrated data-driven framework that synergistically combines Incremental Capacity (IC) analysis, its derivative features, and Differential Voltage (DV) analysis within an advanced ensemble learning architecture to achieve robust and high-fidelity SOH estimation.
The precise monitoring of lithium-ion battery health is paramount. Traditional methods like coulomb counting are simple but suffer from error accumulation and require full charge-discharge cycles. Equivalent circuit models offer a balance but rely heavily on parameter identification and often simplify complex electrochemical aging mechanisms. In recent years, data-driven methods, particularly those leveraging machine learning, have shown great promise. These methods learn the complex mapping between easily measurable battery operational data and its internal state without requiring explicit physical models. Among various diagnostic signals, IC analysis has emerged as a powerful tool. By transforming voltage plateaus in charge curves into identifiable peaks and valleys, the IC curve (dQ/dV vs. V) provides a sensitive fingerprint of internal electrochemical phase transitions. However, reliance on a single signal type can limit robustness against noise and varying aging conditions.
The core innovation of my proposed method lies in a multi-feature fusion strategy. I extract not only standard IC curve features but also compute the Incremental Capacity Slope (IS) curve, which is the first derivative of the IC curve, and the Differential Voltage (DV) curve. This multi-dimensional feature set captures complementary aspects of battery degradation. The IC curve reflects thermodynamic phase equilibria, the IS curve enhances sensitivity to kinetic changes and curve shape, while the DV curve offers an alternative perspective on phase transitions and internal resistance growth. This comprehensive feature vector is then fed into an eXtreme Gradient Boosting (XGBoost) model, a state-of-the-art ensemble learning algorithm known for its high predictive accuracy, speed, and robustness against overfitting.
Methodology: Feature Extraction and Model Framework
The proposed methodology consists of three main stages: (1) Signal Processing and Multi-Feature Extraction, (2) Feature Selection, and (3) Ensemble Learning-based Modeling. The overall technical roadmap is systematic and designed for practical implementation.
1. Signal Processing and Feature Derivation
Starting from the voltage (V) and capacity (Q) data obtained during constant-current charging segments, three key analytical curves are derived.
1.1 Incremental Capacity (IC) Analysis:
The IC curve is calculated as the derivative of capacity with respect to voltage. For discrete data points, it is approximated by:
$$IC = \frac{dQ}{dV} \approx \frac{\Delta Q}{\Delta V}$$
This curve effectively converts voltage plateaus into identifiable peaks, corresponding to major electrochemical phase transitions within the electrode materials of the lithium-ion battery. As the battery ages, these peaks diminish in height and shift along the voltage axis, providing direct indicators of active material loss and lithium inventory depletion.
1.2 Incremental Capacity Slope (IS) Analysis:
To capture the rate of change of the IC curve itself and to obtain features less sensitive to specific sampling intervals, I calculate the IS curve, which is the first derivative of the IC curve:
$$IS = \frac{d(IC)}{dV} = \frac{d^2Q}{dV^2} \approx \frac{\Delta IC}{\Delta V}$$
The IS curve’s inflection points and peaks/valleys offer complementary information to the IC curve, enhancing the model’s ability to discern subtle aging patterns across different lithium-ion battery cells.
1.3 Differential Voltage (DV) Analysis:
The DV curve is the inverse of the IC curve, representing the change in voltage per unit capacity change:
$$DV = \frac{dV}{dQ} \approx \frac{\Delta V}{\Delta Q}$$
Prominent valleys in the DV curve are also linked to electrode phase transitions. The position and depth of these valleys evolve with aging, providing another robust feature set highly correlated with the SOH of the lithium-ion battery.
Raw IC, IS, and DV curves are often noisy. I apply a Gaussian filter for smoothing to facilitate accurate feature point identification without distorting the underlying electrochemical signal.
2. Feature Extraction and Selection
From the smoothed curves, characteristic points are systematically extracted. From the IC curve, I extract the voltage and amplitude of three prominent peaks and two valleys. From the IS curve, I extract the voltage and amplitude of three peaks and three valleys. From the DV curve, I extract the voltage value at its global minimum and the amplitude of its most prominent valley. This results in a broad initial feature pool. To refine this set and reduce dimensionality, I perform Pearson correlation analysis between each feature and the true SOH label. Features with a correlation coefficient magnitude greater than 0.6 are selected for model input. This process yields a final fused feature vector, typically comprising 8 IC features, 7 IS features, and 1 DV feature, as summarized below.
| Curve Type | Feature Points Extracted | Physical Significance | Selected Dimension |
|---|---|---|---|
| IC Curve | Voltage & Amplitude of 3 Peaks (P) and 2 Valleys (V) | Identifies major phase transition states and their evolution. | 8 |
| IS Curve | Voltage & Amplitude of 3 Peaks and 3 Valleys | Captures the rate of change in IC, sensitive to kinetic degradation. | 7 |
| DV Curve | Voltage at Minimum, Amplitude of Key Valley | Provides an alternative view of phase transitions and resistance. | 1 |
| Total Fused Feature Vector Dimension | 16 | ||
3. Ensemble Learning Model: XGBoost Framework
For the regression task of mapping the 16-dimensional feature vector to a SOH value, I employ the XGBoost algorithm. XGBoost is a powerful gradient boosting framework that builds an ensemble of decision trees sequentially, where each new tree corrects the errors of the previous ensemble. Its advantages include handling non-linear relationships effectively, providing feature importance scores, and incorporating regularization to prevent overfitting—a crucial aspect for reliable lithium-ion battery SOH estimation.
The model’s objective function at the t-th iteration is a combination of a loss function \(L\) and a regularization term \(\Omega\):
$$\text{Obj}^{(t)} = \sum_{i=1}^{n} L(y_i, \hat{y}_i^{(t-1)} + f_t(x_i)) + \sum_{k=1}^{t} \Omega(f_k)$$
where \(y_i\) is the true SOH, \(\hat{y}_i^{(t-1)}\) is the prediction from the previous ensemble, and \(f_t(x_i)\) is the new tree being added. Using a second-order Taylor expansion, the objective can be approximated and optimized efficiently. The regularization term for a tree \(f\) is defined as:
$$\Omega(f) = \gamma T + \frac{1}{2} \lambda \sum_{j=1}^{T} w_j^2$$
Here, \(T\) is the number of leaves, \(w_j\) is the score on leaf \(j\), and \(\gamma\) and \(\lambda\) are regularization parameters that control the complexity of the tree. This structured approach allows the XGBoost model to learn a highly accurate and generalizable mapping from the multi-feature space to the lithium-ion battery’s State of Health.
Experimental Validation and Data Acquisition
To validate the proposed method, I conducted accelerated aging tests on commercial lithium iron phosphate (LFP) lithium-ion battery cells. The test setup and procedures were designed to generate realistic aging data.
| Parameter | Specification / Value |
|---|---|
| Cell Chemistry | Lithium Iron Phosphate (LFP) |
| Nominal Capacity | 20 Ah |
| Voltage Range | 2.5 V (Cut-off) to 3.65 V (Cut-off) |
| Charge Protocol | 1C Constant Current (CC) to 3.65V, then Constant Voltage (CV) until current ≤ C/20 |
| Discharge Protocol | 1C Constant Current (CC) to 2.5V |
| Test Temperature | 35°C (Controlled Chamber) |
| Data Sampling Rate | 1 Hz |
Eight cells were cycled continuously using the described protocol. Voltage, current, and accumulated capacity were recorded throughout every cycle. The reference SOH for each cycle was calculated as the ratio of the discharge capacity in that cycle to the initial rated capacity. After data collection, the constant-current charging segments were isolated for analysis. The IC, IS, and DV curves were computed for each cycle, smoothed, and the selected features were extracted to form the dataset. The dataset from eight cells, comprising approximately 6,000 cycles in total, was then partitioned for model development and evaluation.
Results and Discussion
The performance of the proposed multi-feature XGBoost model was rigorously evaluated and compared against several baseline models to isolate the contribution of the feature fusion strategy and the choice of algorithm.
1. Feature Evolution and Model Performance
The extracted features showed strong, systematic correlation with SOH degradation. The voltage positions of the IC peaks exhibited a clear monotonic shift, while their amplitudes decreased. The IS and DV features provided additional, complementary trends. These observable patterns confirmed the suitability of the chosen features for data-driven modeling of lithium-ion battery health.
To evaluate the model, I structured the data such that cells #1, #2, #3, #6, #7, and #8 formed the training set. Cell #5 was used as a validation set for hyperparameter tuning of the XGBoost model (e.g., learning rate, max depth, number of estimators). Cell #4 was held out as a completely independent test set to assess the final model’s generalization capability to an unseen lithium-ion battery.
2. Comparative Analysis of Model Configurations
I constructed six different model configurations to dissect the impact of feature fusion and algorithm choice:
- Models I-III: Use only the combined IC and IS features (15 dimensions) as input, implemented with XGBoost, Support Vector Regression (SVR), and Bagging algorithms, respectively.
- Models IV-VI: Use the fused IC, IS, and DV features (16 dimensions) as input, implemented with the same three algorithms (XGBoost, SVR, Bagging).
The models were assessed using three standard error metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The comparative results on the independent test set are summarized below.
| Model | Input Features | Algorithm | MAE | MAPE (%) | RMSE |
|---|---|---|---|---|---|
| Model I | IC + IS (15 dim) | XGBoost | 0.02235 | 0.1223 | 0.00379 |
| Model II | IC + IS (15 dim) | SVR | 0.01678 | 0.0914 | 0.00149 |
| Model III | IC + IS (15 dim) | Bagging | 0.23905 | 1.3150 | 0.01383 |
| Model IV | IC + IS + DV (16 dim) | XGBoost | 0.01678 | 0.0914 | 0.00149 |
| Model V | IC + IS + DV (16 dim) | SVR | 0.02733 | 0.1492 | 0.00280 |
| Model VI | IC + IS + DV (16 dim) | Bagging | 0.02591 | 0.1418 | 0.00245 |
The results lead to several key findings regarding lithium-ion battery SOH estimation:
1. Superiority of the Fused Feature Approach: Comparing models with the same algorithm but different feature sets clearly demonstrates the benefit of multi-feature fusion. For instance, the XGBoost model with fused features (Model IV, MAE=0.01678) significantly outperforms the XGBoost model with only IC+IS features (Model I, MAE=0.02235), reducing the MAE by approximately 25%. The most dramatic improvement is seen in the SVR algorithm, where adding the DV feature (Model V vs. Model II) reduces the RMSE by over 62%. This confirms that the DV feature carries unique, non-redundant information about lithium-ion battery aging that complements the IC and IS data, leading to a more robust and accurate health estimator.
2. Advantage of the XGBoost Algorithm: Among the three algorithms tested, XGBoost consistently delivered the most stable and accurate performance, especially when combined with the full feature set. While SVR showed low error on the validation set, its performance on the independent test set was more variable and generally worse than XGBoost. The Bagging algorithm showed the highest error in this specific application. XGBoost’s built-in regularization, efficient handling of diverse feature types, and powerful ensemble strategy make it particularly well-suited for modeling the complex, non-linear degradation of lithium-ion batteries.
3. High Estimation Accuracy: The proposed model (Model IV: XGBoost with IC+IS+DV features) achieved excellent accuracy on the independent test set, with an MAE of 0.0167 (1.67% capacity error) and an RMSE of 0.00149. This level of precision is highly competitive and sufficient for practical BMS applications, enabling reliable battery diagnosis, lifetime prediction, and informed decision-making regarding battery use and maintenance.
Conclusion
In this study, I have developed and validated a novel data-driven framework for precise State of Health estimation of lithium-ion batteries. The core contribution is the synergistic fusion of multidimensional health indicators derived from Incremental Capacity (IC), Incremental Capacity Slope (IS), and Differential Voltage (DV) analyses. This multi-feature approach comprehensively captures thermodynamic, kinetic, and resistive aspects of battery degradation, providing a richer and more robust signature of aging than any single-method analysis.
By employing the advanced XGBoost ensemble learning algorithm, the framework effectively learns the complex mapping from these engineered features to the battery’s SOH. Experimental results from accelerated aging tests on commercial LFP cells demonstrate the superiority of this approach. The model incorporating all three feature types consistently outperformed models using only IC and IS features, and the XGBoost algorithm proved to be the most reliable learning engine among those tested.
The proposed method offers a practical, non-invasive, and accurate solution for lithium-ion battery health monitoring. It requires only standard charging data, avoiding the need for specialized electrochemical impedance spectroscopy or destructive tests. The high accuracy and robustness achieved suggest strong potential for implementation in real-world BMS to enhance the safety, longevity, and reliability of lithium-ion battery packs in electric vehicles and energy storage systems. Future work will focus on validating the method under more diverse and dynamic real-world loading profiles and exploring its adaptability across different lithium-ion battery chemistries.