The accurate and real-time estimation of the State of Health (SOH) for lithium-ion batteries remains a paramount challenge within Battery Management Systems (BMS). SOH, typically defined as the ratio of current maximum available capacity to the nominal capacity, is a critical indicator of battery aging, directly impacting the driving range, safety boundaries, and economic value of electric vehicles and energy storage systems. Traditional SOH estimation methods, such as direct capacity measurement through full discharge cycles or electrochemical impedance spectroscopy (EIS), are often impractical for online deployment due to their requirements for long test durations, specific operating conditions, or specialized equipment. Consequently, developing non-invasive, data-driven estimation techniques that leverage operational data has become a key research focus. This article proposes and elaborates on a novel SOH estimation methodology based on the analysis of differential features extracted from the Incremental Capacity (IC) curve during constant-current discharging processes. This method offers a balance between estimation accuracy, computational feasibility, and minimal hardware dependency.
The degradation of a lithium-ion battery is a complex interplay of electrochemical mechanisms, including solid electrolyte interphase (SEI) layer growth, active material loss, lithium inventory depletion, and electrode structural disordering. These internal changes manifest externally in measurable signatures within voltage, current, and temperature profiles. The core innovation of the presented method lies in its focus on the dynamic evolution of the IC curve, which acts as a highly sensitive “electrochemical fingerprint” of the battery’s internal state. By quantitatively tracking changes in specific features of this fingerprint—such as peak positions, heights, and areas—we can establish robust correlations with the overarching capacity fade, enabling precise SOH inference.
Methodology: Incremental Capacity Analysis and Feature Engineering
Incremental Capacity Analysis (ICA) is a potent differential analysis technique that transforms galvanostatic charge/discharge data into a form that amplifies the electrochemical phase transitions within the electrode materials. For a constant-current (CC) process, the IC curve is derived by differentiating capacity (Q) with respect to voltage (V).
$$IC = \frac{dQ}{dV}$$
Since $Q = \int I dt$ and $I$ is constant during a CC phase, the IC value can be practically computed from measured voltage-time data:
$$\frac{dQ}{dV} = I \cdot \frac{dt}{dV} = I \cdot \left(\frac{dV}{dt}\right)^{-1}$$
Each valley or peak in the resulting IC versus voltage ($IC-V$) plot corresponds to a major electrochemical reaction or phase transformation. For instance, in commercial lithium iron phosphate (LFP) cells, the charge/discharge profile is characterized by a flat voltage plateau, which translates into two distinct peaks in the IC domain. These peaks are associated with the two-phase equilibrium between FePO4 and LiFePO4.
As the lithium-ion battery ages, the aforementioned degradation mechanisms cause measurable distortions in the IC curve:
- Peak Voltage Shift ($\Delta V_p$): The voltage at which an IC peak occurs may shift. A positive shift often indicates increased polarization due to growing interfacial impedance (e.g., SEI layer thickening) or increased charge transfer resistance.
- Peak Amplitude Attenuation ($\Delta H_p$): The height of an IC peak diminishes. This is primarily linked to the loss of active lithium inventory and the degradation of active electrode material, reducing the total charge involved in the associated phase transition.
- Peak Area Contraction ($\Delta A_p$): The integrated area under an IC peak decreases. The area is directly proportional to the charge exchanged during that specific electrochemical step. Its contraction is a strong, direct indicator of the loss of usable capacity in the lithium-ion battery.
- Peak Width Variation ($\Delta W_p$): The broadening or narrowing of a peak can reflect changes in the kinetics and homogeneity of the electrode reaction.
Therefore, a feature vector $\mathbf{F}$ can be constructed for each discharge cycle to represent the battery’s health state:
$$\mathbf{F} = [V_{p1}, H_{p1}, A_{p1}, V_{p2}, H_{p2}, A_{p2}, \dots]^T$$
Where the subscripts $p1, p2, …$ denote different IC peaks. The evolution of $\mathbf{F}$ over cycle number $n$ is intrinsically linked to the SOH trajectory. The primary task of the data-driven model is to learn the mapping function $\mathcal{M}$:
$$SOH(n) = \mathcal{M}(\mathbf{F}(n)) + \epsilon$$
where $\epsilon$ represents the model error. This approach requires only the voltage and current data from a standard CC discharge, making it fully compatible with existing BMS sensor suites and ideal for non-invasive, online SOH estimation.
Experimental Data Acquisition and Processing
The validation of this methodology was conducted using cycling data obtained from commercial-grade LFP lithium-ion batteries. The test protocol was designed to simulate realistic aging conditions while collecting high-fidelity data for analysis.
Test Setup: Individual cells were placed in a thermal chamber maintained at a constant temperature of 25°C to minimize the influence of temperature variation on the electrochemical signals. The cells were connected to a programmable battery cycler capable of precise control over charge/discharge currents and voltage limits. Voltage, current, and temperature were sampled at a high frequency (typically 1 Hz or higher) throughout the tests.
Cycling Protocol: The aging test consisted of repeated cycles, each comprising the following steps:
- Rest: The cell was allowed to rest until its temperature stabilized to the ambient chamber temperature (±2°C) for a minimum of 30 minutes to eliminate transient polarization effects.
- Charge: A multi-stage constant current (MCC) charge strategy was employed. The process began with a primary CC charge at 1C rate until the upper voltage cut-off was reached, followed by subsequent CC steps at progressively lower currents until a minimum current threshold was met.
- Rest: A brief rest period followed charging.
- Discharge: A constant-current discharge at a 1C rate was performed until the lower voltage cut-off was reached. This discharge segment provides the $V-t$ data essential for IC curve generation and feature extraction.
The test continued for over 1,200 cycles, with periodic reference performance tests (RPT) to measure the actual capacity. This provided the ground-truth SOH values, calculated as:
$$SOH_{true}(n) = \frac{C_{discharge}(n)}{C_{nominal}} \times 100\%$$
The discharge curves for a representative cell at various cycle stages are shown conceptually below. As expected, the discharge capacity and the average discharge voltage gradually decrease with cycling, visually confirming the capacity and power fade of the lithium-ion battery.
The core data processing flow involves:
- Extracting the CC discharge $V-t$ data for each cycle.
- Calculating the discharged capacity $Q(t)$ by integrating the constant current.
- Computing the IC curve by numerically differentiating $Q$ with respect to $V$. Savitzky-Golay filtering is typically applied to smooth the raw voltage data before differentiation to mitigate noise amplification.
- Identifying and locating the prominent IC peaks (e.g., two peaks for LFP within the 3.2V – 3.4V range).
- Extracting the feature vector $\mathbf{F}$ for each cycle, containing the voltage, height, and area for each identified peak.
Results and Correlation Analysis
A critical step before model training is to analyze the statistical correlation between the extracted IC features and the target SOH. This confirms the relevance of the chosen features and can guide feature selection. The analysis was performed on data from the first several hundred cycles.
The scatter plots and correlation matrix (presented conceptually below) revealed strong and consistent relationships. For the LFP cell, the following observations were made:
- The voltage of the primary IC peaks ($V_{p1}$, $V_{p2}$) showed a significant linear correlation with SOH. The peak voltage generally decreased as SOH declined.
- The peak heights ($H_{p1}$, $H_{p2}$) exhibited an even stronger positive correlation with SOH, decreasing markedly as active material and lithium inventory were depleted.
- The peak areas ($A_{p1}$, $A_{p2}$) also showed a positive correlation, though the relationship could be non-linear, reflecting the direct link to charge loss.
The quantitative correlation coefficients are summarized in the following table:
| Feature | Peak Voltage (V_p) | Peak Height (H_p) | Peak Area (A_p) |
|---|---|---|---|
| Peak 1 | 0.91 | 0.94 | 0.87 |
| Peak 2 | 0.88 | 0.92 | 0.85 |
These high correlation values (all > 0.85) validate that the differential features from the IC curve are indeed potent health indicators for the lithium-ion battery. The temporal evolution of these features over the entire aging test provides a continuous and rich description of the degradation pathway.
SOH Estimation Model Development and Performance
With validated features, the next step is to construct and train a regression model to map the feature vector $\mathbf{F}$ to the SOH value. Several machine learning algorithms were evaluated for this task to identify the most accurate and robust model for the lithium-ion battery SOH estimation.
Model Candidates:
- Artificial Neural Network (ANN): A feedforward network with one or two hidden layers, capable of modeling complex non-linear relationships.
- Gradient Boosting Regressor (GBR): An ensemble of decision trees built sequentially to correct errors, known for high predictive accuracy.
- Support Vector Regression (SVR): A model that finds a hyperplane to fit the data while minimizing error within a tolerance margin.
- Random Forest Regressor (RFR): An ensemble of decision trees built on random data subsets, providing good generalization and feature importance metrics.
The dataset (feature vectors and corresponding true SOH from RPTs) was split into a training set (approximately 70% of cycles) and a testing set (the remaining 30%). The models were trained on the training set and their performance was evaluated on the unseen testing set using standard metrics:
- Root Mean Square Error (RMSE): $RMSE = \sqrt{\frac{1}{N} \sum_{i=1}^{N} (SOH_{true,i} – SOH_{pred,i})^2}$
- Mean Absolute Error (MAE): $MAE = \frac{1}{N} \sum_{i=1}^{N} |SOH_{true,i} – SOH_{pred,i}|$
- Coefficient of Determination (R²): Measures the proportion of variance explained by the model.
The performance comparison is summarized below:
| Model | R² (Test Set) | RMSE (%) | MAE (%) |
|---|---|---|---|
| Artificial Neural Network | 0.994 | 0.43 | 0.31 |
| Gradient Boosting Regressor | 0.990 | 0.58 | 0.45 |
| Random Forest Regressor | 0.985 | 0.71 | 0.55 |
| Support Vector Regression | 0.978 | 0.89 | 0.68 |
The Artificial Neural Network model achieved the best overall performance, with an R² value exceeding 0.99 and an RMSE below 0.5% on the test set. This demonstrates the model’s exceptional ability to capture the complex, non-linear mapping from IC differential features to the SOH of the lithium-ion battery. The Gradient Boosting Regressor also performed very well, offering a strong alternative. A critical long-term validation was performed by applying the trained ANN model to predict SOH at specific high-cycle milestones (e.g., after 400, 800, and 1200 cycles). The predictions remained highly accurate, with the maximum observed error not exceeding 1.2%, confirming the model’s stability and generalization capability over the entire battery lifespan.
Discussion and Mechanism Interpretation
The superior performance of the IC differential feature-based method is not merely a statistical outcome but is rooted in the fundamental electrochemistry of the lithium-ion battery. The IC curve, $dQ/dV$, is mathematically related to the entropic heat flow and the changes in the open-circuit voltage (OCV) curve. The peak features directly reflect the thermodynamics and kinetics of the electrode processes.
The leftward shift of a peak voltage ($\Delta V_p < 0$) can be linked to the increase in internal resistance ($R_i$) and the resulting ohmic polarization during discharge. A simple circuit model illustrates this: the terminal voltage during a CC discharge is $V_{term} = OCV(SOC) – I \cdot R_i$. A growth in $R_i$ causes the voltage to reach a specific OCV plateau (which corresponds to the IC peak) earlier in the discharge, i.e., at a lower apparent terminal voltage. The magnitude of the shift can be approximated by the ohmic drop:
$$\Delta V_p \approx -I \cdot \Delta R_i$$
The decrease in peak area ($\Delta A_p$) is the most direct feature. The area under an IC peak between voltages $V_1$ and $V_2$ is:
$$A_p = \int_{V_1}^{V_2} \frac{dQ}{dV} dV = Q(V_2) – Q(V_1) = \Delta Q_p$$
This $\Delta Q_p$ represents the charge associated exclusively with that electrochemical phase transition. As the lithium-ion battery ages, loss of active lithium (LLI) and loss of active material (LAM) directly reduce the amount of charge that can be exchanged in this reaction, leading to $A_p$ contraction. Therefore, the sum of areas from all major peaks can be seen as a differential proxy for the total usable capacity.
The attenuation of peak height ($\Delta H_p$) is related to the kinetics and homogeneity of the reaction. A lower, broader peak suggests increased overpotential and a wider distribution of reaction sites with varying energies, often caused by electrode particle cracking, inhomogeneous SEI, or localized degradation. This can be conceptually linked to a change in the effective reaction resistance or diffusion coefficients within the lithium-ion battery electrodes.
By monitoring these differential features, the proposed method effectively decouples and quantifies different aging modes (LLI vs. LAM, resistance increase vs. capacity loss) in a non-invasive manner. This provides a much richer diagnostic insight compared to simply tracking the end-of-discharge capacity or the mid-point voltage.
Conclusion and Future Perspectives
This article has detailed a robust, non-invasive, and highly accurate method for estimating the State of Health of lithium-ion batteries. The methodology hinges on extracting and analyzing differential features from the Incremental Capacity curve derived from routine constant-current discharge data. The key IC features—peak voltage, height, and area—serve as sensitive electrochemical fingerprints that quantitatively capture the internal degradation mechanisms of the lithium-ion battery, including lithium inventory loss, active material degradation, and impedance growth.
The experimental validation on LFP cells demonstrated that a machine learning model, particularly an Artificial Neural Network, trained on these features can achieve SOH estimation errors consistently below 1.2% throughout the battery’s cycle life. This performance meets and exceeds the precision requirements for practical BMS applications, enabling reliable range prediction, health-aware charging strategies, and early fault detection without the need for additional hardware or disruptive testing procedures.
The advantages of this approach are multifaceted:
- Non-invasive and Online: It uses only standard BMS sensor data during normal operation.
- High Accuracy and Robustness: Leverages the high sensitivity of differential analysis.
- Diagnostic Insight: The features themselves offer interpretable insights into dominant aging modes.
- Computational Efficiency: Feature extraction and model inference are computationally lightweight, suitable for embedded BMS implementation.
Future research directions will focus on enhancing the generalizability and robustness of this method for the lithium-ion battery:
- Multi-Chemistry and Multi-Cell Validation: Extending the methodology to other prevalent lithium-ion battery chemistries like NMC and LTO, and validating it on battery pack/module data where cell-to-cell variation is significant.
- Feature Adaptation Under Varying Conditions: Developing adaptive algorithms to ensure feature extraction robustness under dynamic loads, varying temperatures, and different state-of-charge windows, as a full discharge may not always be available.
- Integration with Physics-Based Models: Fusing the data-driven IC feature approach with reduced-order electrochemical models to create hybrid models that offer both high accuracy and strong mechanistic interpretability for the lithium-ion battery aging process.
- Lifetime Prediction: Using the trajectory of the extracted differential features not only for current SOH estimation but also for predicting the remaining useful life (RUL) of the lithium-ion battery.
In conclusion, the IC differential feature-based SOH estimation method represents a significant step forward in achieving precise, practical, and insightful battery health management, which is crucial for the safety, longevity, and economic viability of systems powered by lithium-ion batteries.