The accurate estimation of the State of Health (SOH) is critical for ensuring the safety, reliability, and longevity of energy storage lithium battery systems. As an internal state variable, SOH cannot be directly measured by sensors and must be inferred indirectly through the analysis of external operational data. The precision of SOH estimation heavily relies on the quality of extracted health features, which serve as bridges between observable data and the unobservable internal degradation state. Current research faces challenges due to the complexity of internal electrochemical degradation mechanisms and the inability of single features to comprehensively capture the aging process. This article provides a systematic review of feature extraction methodologies for SOH estimation in energy storage lithium batteries, evaluating their physical significance, extraction techniques, and applications.
State of Health and Degradation Mechanisms in Energy Storage Lithium Batteries
The performance degradation of energy storage lithium batteries manifests primarily as capacity fade and impedance rise. Capacity fade reduces the energy storage capability, while impedance growth leads to increased voltage drops and heat dissipation during operation. SOH is commonly defined as the ratio of the current maximum capacity to the nominal capacity, expressed as:
$$ \text{SOH} = \frac{C_{\text{current}}}{C_{\text{nominal}}} \times 100\% $$
Alternatively, SOH can be defined based on internal resistance:
$$ \text{SOH} = \frac{R_{\text{EoL}} – R_{\text{current}}}{R_{\text{EoL}} – R_{\text{new}}} \times 100\% $$
where $R_{\text{EoL}}$ is the resistance at end-of-life, $R_{\text{current}}$ is the current resistance, and $R_{\text{new}}$ is the initial resistance. Underlying these macroscopic phenomena are complex microscopic degradation mechanisms, including Loss of Active Material (LAM) and Loss of Lithium Inventory (LLI). LLI, often caused by solid electrolyte interphase (SEI) growth and lithium plating, directly reduces the available lithium ions for cycling. LAM results from physical and chemical changes in electrode materials, such as particle cracking and structural phase transitions. Ideal health features should correlate explicitly with these mechanisms to provide physically meaningful inputs for SOH estimation.
Feature Extraction Methodologies
Health features are extracted from various operational data, including voltage, current, temperature, and time. These features can be categorized into several groups based on their physical origins and extraction techniques.
Features Based on Voltage-Current Curves
Voltage-current curves, particularly from standard charging protocols like Constant Current-Constant Voltage (CC-CV), offer direct and computationally efficient features. The CC-CV process involves a constant current charge until the voltage reaches a cutoff, followed by a constant voltage phase where current decays. Key features include:
- Time-related features: Constant Current Charging Time (CCCT), Time Interval for Equal Voltage Increase (TEVI), and Time Interval for Equal Current Decrease (TECD). CCCT decreases with aging due to rising internal resistance, while TECD increases as the constant voltage phase prolongs.
- Capacity-related features: Charging capacity within fixed voltage windows. As batteries age, the capacity accumulated in specific voltage ranges diminishes, directly reflecting capacity fade.
- Geometric and statistical features: Slopes, curvatures, and statistical moments (e.g., standard deviation, kurtosis) of voltage/current curves. For example, the area under the voltage curve during charging/discharging can serve as a health indicator.
The correlation between these features and SOH can be quantified using metrics like Pearson Correlation Coefficient (PCC). For instance, CCCT shows a strong negative correlation with SOH, as expressed by:
$$ \text{PCC}(\text{CCCT}, \text{SOH}) < 0 $$
However, a key limitation is the dependency on complete charging cycles, which may not be feasible in real-world applications with partial charging data.
| Feature Type | Description | Correlation with SOH | Limitations |
|---|---|---|---|
| CCCT | Duration of constant current charging phase | Strong negative | Requires full charge cycle |
| TEVI | Time for voltage to increase between fixed points | Negative | Sensitive to initial SOC |
| TECD | Time for current to decay between fixed points in CV phase | Positive | Influenced by temperature |
| ΔV-ΔQ Capacity | Capacity increments in fixed voltage steps | Directly proportional | Noise amplification in differential calculations |
Features Based on Differential Curves
Differential techniques, such as Incremental Capacity Analysis (ICA) and Differential Voltage Analysis (DVA), amplify hidden electrochemical information by transforming charge/discharge data. ICA computes the derivative of capacity with respect to voltage:
$$ \text{IC} = \frac{dQ}{dV} = \frac{I \cdot dt}{dV} $$
where $Q$ is capacity, $V$ is voltage, $I$ is current, and $t$ is time. IC curves exhibit peaks corresponding to electrode phase transitions. Aging causes peak height reduction (indicating LAM), peak area shrinkage, and peak position shifts (reflecting LLI). Similarly, DVA calculates the derivative of voltage with respect to capacity:
$$ \text{DV} = \frac{dV}{dQ} = \frac{dV}{I \cdot dt} $$
DV curves show valleys and peaks that move with degradation. For example, the position of the second inflection point in DV curves shifts leftward as SOH decreases. However, ICA and DVA are sensitive to noise and current rates, necessitating filtering techniques like Savitzky-Golay filters, which may introduce distortion. Recent advances include Virtual ICA (VICA) and Virtual DVA (VDVA), which use machine learning to reconstruct ideal differential curves from arbitrary charging data, enhancing online applicability.
| Feature | Description | Aging Mechanism Association | Challenges |
|---|---|---|---|
| IC Peak Height | Maximum value of dQ/dV | LAM (reduced active material) | Noise sensitivity |
| IC Peak Position | Voltage at peak maximum | LLI (electrode sliding) | Current rate dependence |
| DV Valley Area | Integrated area under DV curve segments | Combined LAM and LLI | Requires low-rate data |
| Peak-to-Peak Distance | Voltage difference between IC peaks | Electrode balancing loss | Overlap at high rates |
Features Based on Hybrid Pulse Power Characterization
Hybrid Pulse Power Characterization (HPPC) tests evaluate the power capability and internal resistance of energy storage lithium batteries across different State of Charge (SOC) levels. An HPPC test sequence includes discharge pulses, rest periods, and charge pulses at fixed SOC intervals. Key features include:
- DC Internal Resistance (DCR): Calculated as $\text{DCR} = \Delta V / \Delta I$ during current pulses. DCR increases with aging due to impedance growth. Accurate DCR estimation requires correction for Open-Circuit Voltage (OCV) changes during pulses.
- OCV-SOC Curve Features: The relationship between OCV and SOC shifts with degradation. Polynomial fitting parameters of the OCV-SOC curve can serve as health indicators. For example, the curve’s slope and intercept changes correlate with SOH.
The HPPC-derived features effectively track impedance rise but depend on precise OCV-SOC modeling. For instance, the DCR at 50% SOC might follow a linear increase with cycling:
$$ \text{DCR}_{50\%} = R_0 + k \cdot N $$
where $R_0$ is initial resistance, $k$ is a degradation rate, and $N$ is cycle count.
Features Based on Electrochemical Impedance Spectroscopy
Electrochemical Impedance Spectroscopy (EIS) provides deep insights into internal processes by applying small AC signals and measuring the impedance response across frequencies. EIS data can be utilized in several ways:
- Raw EIS Data Features: Direct use of impedance magnitude and phase at specific frequencies. Machine learning models can identify critical frequencies highly correlated with SOH, reducing measurement time.
- Equivalent Circuit Model (ECM) Features: Fitting EIS data to ECMs, such as Randles circuits, to extract parameters like charge transfer resistance ($R_{ct}$) and double-layer capacitance ($C_{dl}$). These parameters reflect specific degradation mechanisms, e.g., $R_{ct}$ increase indicates slowed electrode kinetics.
- Distribution of Relaxation Times (DRT) Features: DRT transforms EIS data into time-domain spectra, deconvoluting overlapping processes into distinct peaks. Peak positions and areas in DRT spectra correlate with degradation modes like SEI growth.
For example, the impedance of a typical energy storage lithium battery can be modeled using an ECM consisting of series resistance $R_s$, SEI resistance $R_{SEI}$, and charge transfer resistance $R_{ct}$, with associated capacitances. The total impedance $Z$ is given by:
$$ Z = R_s + \frac{R_{SEI}}{1 + j\omega R_{SEI}C_{SEI}} + \frac{R_{ct}}{1 + j\omega R_{ct}C_{dl}} $$
where $\omega$ is angular frequency. EIS features offer comprehensive diagnostics but require time-intensive measurements and are influenced by temperature and SOC.
| Method | Extracted Features | Advantages | Disadvantages |
|---|---|---|---|
| Raw EIS Data | Impedance at selected frequencies | Avoids model fitting errors | High dimensionality; requires feature selection |
| ECM Fitting | $R_s$, $R_{ct}$, $C_{dl}$, etc. | Clear physical interpretation | Model dependency; fitting instability |
| DRT Analysis | Peak positions, areas in relaxation time spectrum | Decouples overlapping processes | Computationally intensive; noise sensitive |
Multi-Physics Field Features
Beyond electrical signals, multi-physics sensors capture thermal, acoustic, and mechanical changes in energy storage lithium batteries, providing complementary health indicators:
- Thermal Features: Differential Thermal Voltammetry (DTV) calculates the ratio of temperature change to voltage change ($dT/dV$), revealing entropy heat effects during phase transitions. Infrared thermography measures surface temperature distribution; increased temperature variance indicates internal inhomogeneity.
- Acoustic Features: Ultrasonic sensors detect changes in Time-of-Flight (ToF) and signal amplitude due to material property changes, gas formation, or electrode delamination.
- Mechanical Features: Strain gauges measure electrode expansion from lithium plating or gas generation, with strain cycles correlating with SOH degradation.
Entropy-based features, such as Sample Entropy ($S_{\text{sample}}$) of voltage signals, quantify the increasing disorder with aging:
$$ S_{\text{sample}} = -\log \frac{A}{B} $$
where $A$ and $B$ are probabilities of similar sequences in the time series. Multi-physics features enhance robustness but require additional sensors and data fusion algorithms.
Public Datasets for Energy Storage Lithium Battery Research
Standardized public datasets are essential for developing and validating SOH estimation algorithms. Key datasets include:
| Dataset | Battery Types | Testing Conditions | Key Parameters |
|---|---|---|---|
| NASA | 18650 LCO cells | Various temperatures; cycling to 70-80% SOH | Voltage, current, temperature, EIS |
| CALCE | LCO, LFP, NMC formats | Multiple temperatures and SOC levels; dynamic profiles | Capacity, impedance, OCV |
| Oxford | Kokam pouch cells | 40°C; Artemis driving cycles | Drive cycle data, characterization tests |
| XJTU | 18650 NMC cells | Multiple charging strategies; 1 Hz sampling | Full lifecycle voltage, current, temperature |
| HUST | 18650 LFP cells | 30°C; varied discharge protocols | Cycle life data under diverse loads |
These datasets facilitate benchmarking of feature extraction methods and machine learning models under controlled and real-world conditions, accelerating advancements in energy storage lithium battery management.
Summary and Future Perspectives
Feature extraction is a cornerstone of accurate SOH estimation for energy storage lithium batteries. This review has detailed methodologies ranging from voltage-current curves to multi-physics approaches, highlighting their physical bases and practical challenges. Key insights include:
- Voltage-current features are efficient but often require complete cycles.
- ICA/DVA offer mechanistic insights but are noise-sensitive and rate-dependent.
- HPPC and EIS features effectively track impedance and degradation mechanisms but need precise measurements and modeling.
- Multi-physics features provide additional dimensions but increase system complexity.
Despite progress, no single feature reliably ensures robust SOH estimation under variable real-world conditions. Future research should focus on:
- Standardization and Benchmarking: Establishing unified evaluation protocols and public datasets to enable objective comparisons and accelerate technology transfer for energy storage lithium battery systems.
- Multi-Feature Fusion: Integrating electrical, thermal, acoustic, and mechanical features through advanced algorithms (e.g., deep learning) to create comprehensive health indicators that overcome the limitations of individual features.
- Physics-Informed Data-Driven Models: Combining physical models (e.g., electrochemical or equivalent circuit models) with machine learning, such as Physics-Informed Neural Networks (PINNs), to enhance interpretability, data efficiency, and generalization. For example, incorporating degradation dynamics into loss functions:
$$ \mathcal{L} = \mathcal{L}_{\text{data}} + \lambda \mathcal{L}_{\text{physics}} $$
where $\mathcal{L}_{\text{data}}$ is the prediction error, $\mathcal{L}_{\text{physics}}$ enforces physical constraints, and $\lambda$ is a weighting factor. This approach can improve SOH estimation with limited data and ensure consistency with fundamental principles.
In conclusion, the evolution of feature extraction techniques will play a pivotal role in advancing the reliability and sustainability of energy storage lithium batteries, supporting their applications in electric vehicles, grid storage, and circular economy initiatives.