In this review, I examine the state of health (SoH) estimation methods that are most relevant to retired lithium ion battery packs. The global energy transition has pushed electric vehicle production into a period of rapid expansion, and the resulting wave of retired lithium ion battery systems is becoming a major material, economic, and environmental issue. When a lithium ion battery reaches the end of its first life, it still often retains between 60% and 80% of its initial capacity, which makes second-life use attractive for telecommunication base stations, distributed microgrids, low-speed electric vehicles, backup power, and stationary storage. The central problem is that the remaining value of a retired lithium ion battery cannot be realized safely without a reliable SoH estimate. A precise SoH estimate is the gatekeeper for sorting, grading, pricing, reuse, and final recycling.
I define SoH in several complementary ways because a lithium ion battery ages through capacity fade, resistance rise, power fade, thermal instability, and material degradation. The most common capacity-based definition is:
$$ SoH_C = \frac{C_{current}}{C_{rated}} \times 100\% $$
where \(C_{current}\) is the present usable capacity and \(C_{rated}\) is the rated capacity. A resistance-based definition is also important for power-sensitive second-life applications:
$$ SoH_R = \frac{R_{EOL} – R_{current}}{R_{EOL} – R_{new}} \times 100\% $$
where \(R_{new}\) is the fresh-cell resistance, \(R_{current}\) is the present resistance, and \(R_{EOL}\) is the end-of-life resistance threshold. Because a retired lithium ion battery can have good capacity but poor power capability, I prefer a multi-indicator definition:
$$ SoH_{multi} = w_C SoH_C + w_R SoH_R + w_E SoH_E + w_T SoH_T $$
with weights \(w_C + w_R + w_E + w_T = 1\). The terms \(SoH_E\) and \(SoH_T\) can represent energy efficiency and thermal health, respectively. This multi-indicator view is essential when a lithium ion battery is evaluated for a new duty cycle that differs from its original vehicle duty cycle.
The technical challenge is not simply to measure one number. I see three coupled difficulties. First, retired lithium ion battery cells often lack complete historical data, so empirical models trained on new cells may fail. Second, aging under complex vehicle and storage conditions is a multi-factor coupling process involving temperature, state of charge (SOC), current rate, depth of discharge, calendar time, and mechanical stress. Third, second-life sorting requires fast, low-cost, and scalable detection, while laboratory-grade methods such as full capacity tests are slow and sometimes destructive to the remaining life of the lithium ion battery.
Table 1 summarizes the main families of SoH estimation methods that I compare in this review.
| Method family | Core principle | Typical inputs | Typical outputs | Main strength | Main limitation | Time scale |
|---|---|---|---|---|---|---|
| Direct measurement | Measure physical quantities linked to aging | Current, voltage, temperature, impedance | Capacity, resistance, impedance parameters | High physical interpretability | Slow or sensitive to test conditions | Minutes to hours |
| Model-based | Fit a mathematical or physical model and infer SoH | Voltage response, current, temperature | Model parameters, SoH | Good for online estimation | Model mismatch and parameter drift | Milliseconds to seconds |
| Data-driven | Learn mapping from data to SoH | Voltage, current, temperature, impedance, features | SoH or degradation class | Handles nonlinearity and complex conditions | Needs data and may lack interpretability | Milliseconds to seconds |
| Hybrid | Fuse physical models with data learning | All available measurements | Robust SoH and uncertainty | Balances accuracy and generalization | Higher implementation complexity | Milliseconds to seconds |
Direct Measurement Methods
I begin with direct measurement because these methods provide the physical reference against which other methods are often validated. For a retired lithium ion battery, direct measurement can be performed on cells, modules, or packs, but the interpretation changes with scale. At the pack level, cell inconsistency, contact resistance, and thermal gradients can obscure the true aging state of individual cells.
Capacity Test Method
The capacity test method charges and discharges a lithium ion battery according to a standard protocol and computes the usable capacity by Coulomb integration:
$$ Q = \int_{t_0}^{t_1} I(t) \, dt $$
The SoH is then obtained from:
$$ SoH_C = \frac{Q_{discharge}}{Q_{nominal}} \times 100\% $$
In practice, I must account for Coulombic efficiency, temperature dependence, and rate dependence. An enhanced Coulomb counting model can be written as:
$$ Q_{corrected} = \int_{t_0}^{t_1} \eta(T, I, SOC) I(t) \, dt + \Delta Q_{history} $$
where \(\eta(T, I, SOC)\) is a dynamic efficiency factor and \(\Delta Q_{history}\) is a history-based correction term. This correction can improve capacity estimation to roughly \(\pm 1.5\%\) under controlled conditions, but the method still requires a full charge and discharge cycle. The full test may take four to eight hours, and repeated deep cycling accelerates aging of the retired lithium ion battery. Therefore, I do not regard the capacity test as a practical primary method for high-throughput second-life sorting, although it remains a valuable benchmark.
| Capacity test variant | Protocol | Estimated quantity | Reported accuracy | Main drawback |
|---|---|---|---|---|
| Standard full charge-discharge | Constant current-constant voltage, then constant current discharge | Usable capacity | High under laboratory conditions | Long duration and aging side effects |
| Partial cycle extrapolation | Partial charge or discharge segments | Capacity estimate | Moderate | Requires voltage window selection |
| Dynamic efficiency Coulomb counting | Variable current with efficiency compensation | Corrected capacity | About \(\pm 1.5\%\) | Needs parameter calibration |
| Incremental capacity analysis | Differential capacity versus voltage | Degradation modes | Mode-sensitive | Noise amplification in derivatives |
Internal Resistance Test Method
Internal resistance is a strong indicator of power capability and aging in a lithium ion battery. I often use the hybrid pulse power characterization (HPPC) method because it can extract ohmic and polarization resistance in a short test. The simplest ohmic resistance estimate is:
$$ R_0 = \frac{\Delta V}{\Delta I} $$
where \(\Delta V\) is the instantaneous voltage change and \(\Delta I\) is the current pulse amplitude. A more complete polarization resistance can be expressed as:
$$ R_{pol} = \frac{V_{relax} – V_{pulse}}{I_{pulse}} $$
For a lithium ion battery, the total impedance rise is linked to several aging modes:
$$ R_{total} = R_{ohmic} + R_{SEI} + R_{ct} + R_{diff} $$
where \(R_{SEI}\) is the solid electrolyte interphase resistance, \(R_{ct}\) is the charge-transfer resistance, and \(R_{diff}\) is the diffusion-related resistance. HPPC is widely used because it can be completed in minutes, but its pulse duration is usually on the order of seconds, which may miss fast dynamic responses. Short-pulse techniques using microsecond-level excitation can suppress concentration polarization and reveal the intrinsic response of the lithium ion battery. Multi-frequency or superimposed sinusoidal excitation can further improve robustness, especially when the solid electrolyte interphase thickens in later aging stages.
| Resistance method | Excitation | Extracted parameter | Strength | Weakness |
|---|---|---|---|---|
| Direct current pulse | Second-level pulse | Ohmic resistance | Simple and industrial | Limited dynamic information |
| Short pulse | Microsecond-level pulse | Intrinsic resistance | Suppresses concentration polarization | Requires fast hardware |
| Multi-frequency excitation | Superimposed sinusoids | Frequency-resolved resistance | Robust to internal uncertainty | More complex signal processing |
| HPPC | Pulse sequence | Ohmic and polarization resistance | Standardized for power assessment | Test time and SOC dependence |
Impedance Spectroscopy Analysis
Electrochemical impedance spectroscopy (EIS) provides a frequency-domain view of a lithium ion battery. A small sinusoidal perturbation is applied over a wide frequency range, and the complex impedance is measured:
$$ Z(\omega) = \frac{V(\omega)}{I(\omega)} = Z'(\omega) + j Z”(\omega) $$
An equivalent circuit model can then be fitted to the impedance spectrum. A common form is:
$$ Z_{model}(\omega) = R_0 + \frac{R_{ct}}{1 + j \omega R_{ct} C_{dl}} + \frac{Z_W(\omega)}{1 + j \omega R_{ct} C_{dl}} $$
where \(R_0\) is the ohmic resistance, \(R_{ct}\) is the charge-transfer resistance, \(C_{dl}\) is the double-layer capacitance, and \(Z_W\) is the Warburg diffusion impedance. For a retired lithium ion battery, EIS can reveal changes in the solid electrolyte interphase, charge-transfer kinetics, and diffusion limitations. In the high-frequency region above about 1 kHz, the ohmic resistance often shows a linear relationship with capacity fade, for example:
$$ R_0 = a + b(1 – SoH) $$
The phase angle at a selected frequency can also be correlated with aging:
$$ \phi(f^*) = c + d(1 – SoH) $$
Despite its precision, conventional EIS has practical drawbacks. A full-frequency scan can take minutes, which is too slow for inline sorting of a retired lithium ion battery. Parameter extraction depends on equivalent circuit selection and can be disturbed by noise. Operando EIS, which measures impedance during charging or discharging, offers a promising route to online degradation monitoring, but it still requires careful thermal and electrical control.
| Frequency band | Physical interpretation | Aging indicator | Use in retired lithium ion battery assessment |
|---|---|---|---|
| High frequency, above 1 kHz | Ohmic and contact processes | Ohmic resistance rise | Fast screening of connection and electrolyte resistance |
| Mid frequency, 1 Hz to 1 kHz | Solid electrolyte interphase and charge transfer | Interfacial resistance rise | Detection of SEI growth and kinetics degradation |
| Low frequency, below 1 Hz | Diffusion and accumulation | Warburg coefficient change | Assessment of lithium inventory and diffusion limitation |
| Phase-angle point | Frequency-dependent dynamics | Nonlinear aging correlation | Feature for compact SoH models |
Model-Based Methods
Model-based methods attempt to represent the internal behavior of a lithium ion battery through mathematical or physical equations. I group them into equivalent circuit models, electrochemical models, and mathematical models. The general workflow is model construction, parameter identification, and state estimation.
$$ \text{Model construction} \rightarrow \text{Parameter identification} \rightarrow \text{SoH estimation} $$
Equivalent Circuit Model Method
An equivalent circuit model (ECM) uses resistors, capacitors, and voltage sources to reproduce the terminal voltage response of a lithium ion battery. The simplest Rint model is:
$$ V_t = OCV(SOC) – I R_0 $$
The Thevenin model adds one RC branch:
$$ V_t = OCV(SOC) – I R_0 – V_1 $$
$$ \dot{V}_1 = -\frac{V_1}{R_1 C_1} + \frac{I}{C_1} $$
More complex ECMs include two or more RC branches to capture charge transfer and diffusion. The parameters of an ECM change with aging. For example, I can express empirical mappings as:
$$ R_0(SoH) = R_{0,new} \exp[\alpha (1 – SoH)] $$
$$ R_1(SoH) = R_{1,new} \exp[\beta (1 – SoH)] $$
$$ C_1(SoH) = C_{1,new} [1 – \gamma (1 – SoH)] $$
These mappings show that resistance terms often grow exponentially while capacitance terms may decline linearly. Once the model is identified, a state estimator such as a Kalman filter can track the parameters and the SoH. A discrete state-space form is:
$$ x_{k+1} = A x_k + B u_k + w_k $$
$$ y_k = C x_k + D u_k + v_k $$
where \(w_k\) and \(v_k\) are process and measurement noise. Adaptive versions of the Kalman filter, including unscented and square-root implementations, can improve robustness for a retired lithium ion battery under dynamic load. Auto-regressive ECMs can jointly identify ohmic resistance and capacity from dynamic charge-discharge data, which is attractive for online second-life assessment.
| ECM type | Circuit structure | Advantages | Limitations | Typical use |
|---|---|---|---|---|
| Rint | Voltage source and ohmic resistance | Simple and fast | Ignores polarization | Initial screening |
| RC network | Ohmic resistance plus one or more RC branches | Captures polarization | Parameter coupling | Online SoH estimation |
| PNGV | Ohmic resistance, polarization branch, and capacitance | Better dynamic response | More parameters | Automotive battery management |
| Thevenin | Voltage source, series resistance, RC branch | Good balance of accuracy and complexity | SOC and temperature dependence | Kalman-filter-based SoH |
| Nonlinear ECM | Nonlinear voltage-current elements | Handles nonlinearity | Identification difficulty | Wide operating range |
Electrochemical Model Method
Electrochemical models describe the internal reaction and transport processes of a lithium ion battery using partial differential equations. The pseudo-two-dimensional (P2D) model couples solid-phase diffusion, electrolyte transport, and charge conservation. A representative solid-phase diffusion equation is:
$$ \frac{\partial c_s}{\partial t} = \frac{D_s}{r^2} \frac{\partial}{\partial r} \left( r^2 \frac{\partial c_s}{\partial r} \right) $$
The electrolyte concentration can be described by:
$$ \frac{\partial c_e}{\partial t} = \nabla \cdot (D_e \nabla c_e) + \frac{a_s j}{F} $$
The Butler-Volmer equation relates current density to overpotential:
$$ j = i_0 \left[ \exp \left( \frac{\alpha_a F \eta}{RT} \right) – \exp \left( – \frac{\alpha_c F \eta}{RT} \right) \right] $$
These equations allow me to interpret capacity fade as lithium inventory loss, active material loss, and impedance rise. The single-particle model (SPM) simplifies the P2D model by assuming uniform spherical particles and neglecting electrolyte concentration gradients:
$$ \frac{\partial c_s}{\partial t} = D_s \frac{1}{r^2} \frac{\partial}{\partial r} \left( r^2 \frac{\partial c_s}{\partial r} \right) $$
SPM is computationally much faster than P2D, often by several orders of magnitude, but it may miss electrolyte-limited behavior at high rates. Multi-scale models combine electrochemical, thermal, and mechanical degradation. For a retired lithium ion battery, electrochemical models are powerful for diagnosis, but they are difficult to parameterize quickly because many internal states are not directly measurable.
| Model | Physical detail | Computation cost | Strength | Limitation |
|---|---|---|---|---|
| P2D | High: solid diffusion, electrolyte transport, charge conservation | High | Accurate internal state simulation | Difficult parameter identification |
| SPM | Medium: solid diffusion only | Low to medium | Fast and physically meaningful | Neglects electrolyte gradients |
| Multi-scale coupled | Very high: electrochemical, thermal, mechanical | Very high | Captures coupled degradation | Complex calibration |
| Reduced-order electrochemical | Medium: simplified transport | Medium | Better online feasibility | Reduced accuracy in extreme conditions |
Mathematical Model Method
Mathematical models simplify aging into empirical or probabilistic relationships. I use them when the available data are limited and a fast estimate is needed. An empirical regression for a lithium ion battery may take the form:
$$ SoH = \alpha_0 + \alpha_1 T + \alpha_2 C_{rate} + \alpha_3 SOC + \alpha_4 t + \alpha_5 T \cdot C_{rate} + \cdots $$
A Markov chain can represent transitions among health states:
$$ P(X_{t+1} = j \mid X_t = i) = p_{ij} $$
where the states may be healthy, mildly aged, and severely aged. Bayesian inference can quantify uncertainty:
$$ p(SoH \mid D) \propto p(D \mid SoH) p(SoH) $$
For a battery pack, Copula functions can model dependence among cells:
$$ F_{SoH_1, \ldots, SoH_n}(x_1, \ldots, x_n) = C(F_1(x_1), \ldots, F_n(x_n)) $$
These methods are computationally efficient and can be deployed quickly, but they depend heavily on the training or calibration population. If a retired lithium ion battery has a different chemistry, format, or duty history, the mathematical model may need re-identification.
| Mathematical method | Form | Use case | Strength | Weakness |
|---|---|---|---|---|
| Empirical regression | Explicit algebraic equation | Fast screening | Simple and transparent | Limited extrapolation |
| Markov chain | State transition probabilities | Calendar aging and staged degradation | Captures discrete transitions | State definition is subjective |
| Bayesian inference | Posterior distribution | Uncertainty quantification | Provides confidence bounds | Requires prior knowledge |
| Copula model | Joint distribution of cells | Pack-level sorting | Models cell correlation | Complex for large packs |
Data-Driven Methods
Data-driven methods learn the relationship between measured signals and SoH without requiring an explicit physical model. For a retired lithium ion battery, this is attractive because historical data may be incomplete and because aging patterns are nonlinear. I divide data-driven methods into neural networks, fuzzy logic, and regression algorithms.
Neural Network Method
A feedforward neural network builds a nonlinear mapping from input features to SoH:
$$ y = \sigma(W_L \cdots \sigma(W_1 x + b_1) \cdots + b_L) $$
The weights are updated by gradient descent:
$$ W_{ij}^{(l)} \leftarrow W_{ij}^{(l)} – \eta \frac{\partial L}{\partial W_{ij}^{(l)}} $$
Recurrent networks are better suited to time-series data from a lithium ion battery. A long short-term memory (LSTM) cell uses gates to control information flow:
$$ f_t = \sigma(W_f [h_{t-1}, x_t] + b_f) $$
$$ i_t = \sigma(W_i [h_{t-1}, x_t] + b_i) $$
$$ \tilde{C}_t = \tanh(W_C [h_{t-1}, x_t] + b_C) $$
$$ C_t = f_t C_{t-1} + i_t \tilde{C}_t $$
$$ o_t = \sigma(W_o [h_{t-1}, x_t] + b_o) $$
$$ h_t = o_t \tanh(C_t) $$
A bidirectional LSTM processes the sequence in both directions:
$$ \overrightarrow{h_t} = LSTM_f(x_t, \overrightarrow{h_{t-1}}) $$
$$ \overleftarrow{h_t} = LSTM_b(x_t, \overleftarrow{h_{t+1}}) $$
$$ y_t = g(\overrightarrow{h_t}, \overleftarrow{h_t}) $$
For a retired lithium ion battery, incremental capacity analysis can provide health features such as voltage plateau location and peak height. These features can be fed into a bidirectional LSTM to capture temporal dependence. Nonlinear autoregressive networks can also map health indicators directly to SoH. The main advantage is automatic feature extraction; the main disadvantage is the need for large, representative, and labeled datasets.
| Neural architecture | Input type | Strength | Weakness | Retired lithium ion battery relevance |
|---|---|---|---|---|
| Feedforward network | Static features | Simple and fast | Ignores sequence order | Quick grading from extracted features |
| Recurrent network | Time series | Captures temporal dependence | Vanishing gradients | Online voltage and current tracking |
| LSTM | Time series | Long-range memory | More parameters | Dynamic load and aging history |
| Bi-LSTM | Bidirectional sequence | Uses past and future context | Not fully causal for online use | Offline feature-rich estimation |
| Autoencoder | High-dimensional data | Feature compression | Requires training data | Denoising and anomaly detection |
Fuzzy Logic Method
Fuzzy logic handles uncertainty and linguistic rules. I can define a membership function as:
$$ \mu_A(x): X \rightarrow [0,1] $$
A fuzzy rule may be written as:
$$ R_i: \text{IF } x_1 \text{ is } A_{i1} \text{ AND } \cdots \text{ THEN } SoH \text{ is } B_i $$
The defuzzified output can be computed by weighted averaging:
$$ SoH = \frac{\sum_i w_i c_i}{\sum_i w_i} $$
For a lithium ion battery, the inputs can include charge-transfer resistance, ohmic resistance, temperature, and capacity fade rate. Fuzzy rules such as “low impedance and low fade imply high health” or “high impedance and high fade imply low health” can be tuned for second-life sorting. Fuzzy logic is robust to parameter drift and measurement noise, but rule design requires expert knowledge.
| Fuzzy input | Linguistic terms | Output | Advantage | Risk |
|---|---|---|---|---|
| Ohmic resistance | Low, medium, high | SoH class | Easy to interpret | Threshold selection |
| Charge-transfer resistance | Small, moderate, large | Power health | Captures kinetics | EIS dependency |
| Capacity fade rate | Slow, moderate, fast | Remaining life | Uses historical trend | Requires history |
| Temperature | Low, normal, high | Operating risk | Thermal awareness | Local sensor bias |
Regression Algorithms
Regression methods establish a statistical mapping from features to SoH. Gaussian process regression (GPR) is a Bayesian non-parametric method:
$$ f(x) \sim GP(m(x), k(x, x’)) $$
A common kernel is the squared exponential:
$$ k(x, x’) = \sigma_f^2 \exp \left( -\frac{(x – x’)^2}{2l^2} \right) $$
The predictive distribution is:
$$ p(y_* \mid X, y, x_*) = \mathcal{N}(\mu_*, \sigma_*^2) $$
Support vector regression solves:
$$ \min_{w,b} \frac{1}{2} \|w\|^2 + C \sum_i (\xi_i + \xi_i^*) $$
Random forest averages many decision trees:
$$ \hat{y} = \frac{1}{B} \sum_{b=1}^{B} T_b(x) $$
For a retired lithium ion battery, I can extract features such as capacity fade rate, impedance growth rate, voltage curve statistics, and temperature exposure. Ensemble models often outperform single models because they combine different inductive biases. A hybrid feature-learning pipeline may first use a convolutional network to extract time-frequency health indicators, then use a random forest or GPR to estimate SoH. Transfer learning and Bayesian neural networks can mitigate the small-sample problem when only a few retired lithium ion battery cells are available for a new chemistry.
| Regression method | Model form | Uncertainty | Data demand | Interpretability |
|---|---|---|---|---|
| Linear regression | Linear coefficients | Low | Low | High |
| Support vector regression | Kernel margin | Low to medium | Medium | Medium |
| Gaussian process regression | Bayesian kernel | High | Medium | Medium |
| Random forest | Tree ensemble | Medium | Medium to high | Medium |
| Hybrid ensemble | Multiple learners | Medium to high | High | Low to medium |
Hybrid and Multi-Modal Estimation
I find that the strongest practical direction is not to choose one family but to fuse them. A direct measurement gives a physical anchor, a model gives continuity, and a data-driven learner gives adaptability. A generic fusion form is:
$$ SoH = \Phi(SoH_{direct}, SoH_{model}, SoH_{data}) $$
A weighted fusion can be optimized by least squares:
$$ w^* = \arg\min_w \left\| y – \sum_i w_i \hat{y}_i \right\|^2 + \lambda \|w\|^2 $$
For a retired lithium ion battery, I can use EIS-derived impedance features as inputs to an ECM, then use a neural network to correct the residual between the ECM output and the measured capacity. This physics-informed residual learning can reduce the data requirement while preserving physical interpretability. Another option is a Bayesian fusion framework:
$$ p(SoH \mid y_1, y_2, \ldots, y_n) \propto \prod_i p(y_i \mid SoH) p(SoH) $$
Table 10 summarizes common fusion strategies.
| Fusion strategy | Structure | Strength | Challenge |
|---|---|---|---|
| Direct plus model | Use measured resistance to initialize ECM | Physical consistency | Measurement time |
| Model plus data | ECM residual corrected by machine learning | Improved accuracy | Training data alignment |
| Feature-level fusion | Combine EIS, voltage, and thermal features | Rich information | Feature synchronization |
| Decision-level fusion | Weighted voting or stacking | Robust to single-model failure | Weight optimization |
| Physics-informed learning | Add physical loss to neural network | Better extrapolation | Complex loss design |
Challenges in Retired Lithium Ion Battery SoH Estimation
I identify three core challenges that remain unresolved for large-scale second-life deployment. The first is the efficiency-accuracy contradiction. A full capacity test is accurate but slow and may degrade the lithium ion battery further. The second is limited generalization. An ECM or empirical model calibrated on one lithium ion battery chemistry may not transfer to another chemistry, format, or aging path. The third is the black-box and small-sample problem in data-driven methods. Deep learning may achieve low error on a training set, but it may fail when the retired lithium ion battery has a different history or when labeled data are scarce.
| Challenge | Root cause | Consequence | Possible mitigation |
|---|---|---|---|
| Efficiency-accuracy trade-off | Full tests are slow; fast tests are indirect | Sorting bottleneck | Short-pulse resistance, partial charging, EIS feature selection |
| Poor generalization | Model parameters depend on chemistry and history | High error on new retired packs | Transfer learning, adaptive parameters, multi-chemistry training |
| Data scarcity | Limited labeled retired lithium ion battery data | Overfitting | Data augmentation, physics-informed learning, generative models |
| Cell inconsistency | Pack aging is non-uniform | Pack SoH hides weak cells | Cell-level sensing, correlation models, weakest-cell estimation |
| Safety uncertainty | Internal short and thermal risk are not directly observed | Unsafe reuse | Multi-sensor fusion, impedance monitoring, thermal imaging |
For pack-level estimation, I must also consider inconsistency. A pack SoH is not simply the average of cell SoH values. If one lithium ion battery cell ages faster, the usable pack capacity may be limited by the weakest cell:
$$ C_{pack,usable} \approx \min_i C_i $$
Series and parallel connections further complicate the mapping from cell health to pack health. Therefore, I argue that retired lithium ion battery assessment should combine cell-level features, module-level statistics, and pack-level constraints.
Future Trends
I see several research directions that can move SoH estimation from laboratory accuracy to industrial scalability.
Multi-Physics Coupling Modeling
Aging in a lithium ion battery is electrochemical, thermal, and mechanical at the same time. A coupled thermal model can be written as:
$$ \frac{\partial T}{\partial t} = \alpha \nabla^2 T + \frac{Q_{gen}}{\rho C_p} $$
Mechanical strain can be represented as:
$$ \epsilon = \frac{\Delta L}{L_0} = f(SOC, SoH, T) $$
By coupling these fields, I can better predict capacity fade and resistance rise under complex second-life duty cycles. The goal is a reduced-order multi-physics model that is accurate enough for diagnosis and fast enough for online use.
Intelligent Sensing
New sensors can provide direct or indirect aging signatures. Impedance-temperature composite sensors, surface acoustic wave sensors, and fiber-optic strain sensors can be embedded in a lithium ion battery module. These sensors can capture local temperature, pressure, and strain, which are often missed by conventional voltage and current measurements. The resulting data streams can feed a fusion estimator and improve early fault detection.
Algorithm Architecture Optimization
Transfer learning, meta-learning, and reinforcement learning can reduce the data burden. A physics-informed loss can regularize a neural network:
$$ \mathcal{L}_{total} = \mathcal{L}_{data} + \lambda \mathcal{L}_{physics} $$
For example, the physical loss can penalize deviations from a diffusion equation:
$$ \mathcal{L}_{physics} = \left\| \frac{\partial c_s}{\partial t} – D_s \nabla^2 c_s \right\|^2 $$
Such a model can learn from fewer labeled samples while respecting the known behavior of a lithium ion battery. Edge computing can also bring SoH estimation directly to the battery management system, reducing communication latency and enabling real-time sorting.
| Future direction | Key idea | Expected benefit | Implementation barrier |
|---|---|---|---|
| Multi-physics coupling | Electrochemical, thermal, and mechanical coupling | Better degradation prediction | High model complexity |
| Intelligent sensing | Embedded impedance, temperature, and strain sensors | Rich local data | Cost and integration |
| Physics-informed learning | Combine data loss with physical constraints | Improved generalization | Loss balancing |
| Transfer learning | Reuse knowledge across chemistries and packs | Less labeled data | Domain shift |
| Digital twin | Virtual replica of each retired lithium ion battery | Predictive sorting and lifetime planning | Model calibration |
| Edge intelligence | On-device estimation and decision-making | Low latency and privacy | Computing power limits |
Data-Physics Synergy
I strongly favor a data-physics synergy because neither pure data nor pure physics is sufficient. Pure data-driven models can fit complex patterns but may violate conservation laws. Pure physical models are interpretable but difficult to parameterize for every retired lithium ion battery. A hybrid model can use physical equations to generate synthetic data, and then use machine learning to correct residual errors. A general hybrid estimator can be written as:
$$ \hat{SoH} = f_{physics}(x; \theta_p) + f_{data}(x; \theta_d) $$
where \(\theta_p\) are physical parameters and \(\theta_d\) are data-driven parameters. Joint training minimizes:
$$ \min_{\theta_p, \theta_d} \sum_{k} \left( SoH_k – \hat{SoH}_k \right)^2 + \lambda_1 \|\theta_p\|_1 + \lambda_2 \|\theta_d\|_2^2 $$
This formulation allows the physical part to extrapolate and the data part to adapt. For retired lithium ion battery sorting, such a model can be calibrated on a small number of full tests and then applied to many partially tested packs.
Practical Estimation Workflow
I propose a practical workflow for second-life assessment. The first stage is a safety check: visual inspection, insulation resistance, open-circuit voltage, and temperature stability. The second stage is a fast feature acquisition: short-pulse resistance, partial charge or discharge, and perhaps a limited-frequency EIS scan. The third stage is feature extraction: capacity-related features, resistance-related features, and thermal features. The fourth stage is SoH estimation using a hybrid model. The fifth stage is grading and decision-making: reuse, remanufacture, or recycle. Table 12 summarizes this workflow.
| Stage | Action | Measured or derived quantity | Decision |
|---|---|---|---|
| Safety screening | Visual, insulation, open-circuit voltage, temperature | Safety flags | Reject unsafe lithium ion battery units |
| Fast characterization | Short pulse, partial cycle, limited EIS | Resistance, voltage response, impedance | Generate health features |
| Feature engineering | Extract capacity, resistance, thermal, and consistency features | Feature vector | Prepare model input |
| SoH estimation | Hybrid model inference | SoH and uncertainty | Grade the retired lithium ion battery |
| Second-life planning | Match SoH to application requirements | Remaining life projection | Reuse, remanufacture, or recycle |
I emphasize that uncertainty must be reported alongside SoH. A point estimate without confidence bounds is not enough for a retired lithium ion battery because safety and warranty decisions depend on risk. Bayesian methods, ensemble variance, and conformal prediction can provide uncertainty intervals. For example, a Gaussian process can output:
$$ SoH \sim \mathcal{N}(\mu(x), \sigma^2(x)) $$
A conservative grading rule may use the lower confidence bound:
$$ SoH_{safe} = \mu(x) – z_{\alpha} \sigma(x) $$
where \(z_{\alpha}\) is selected for the desired confidence level. This is especially important when a lithium ion battery is reused in a stationary application with limited thermal management.
Comparative Discussion
In my assessment, direct measurement methods remain the reference for accuracy, but they are not scalable for every retired lithium ion battery. Capacity tests are too slow, and full EIS is too complex for high-throughput lines. Resistance tests are faster, but they only capture part of the aging state. Model-based methods provide continuity and physical meaning, but they need careful parameter identification. Data-driven methods provide flexibility, but they need data and may not extrapolate. Hybrid methods offer the best compromise, especially when they combine fast impedance features, a reduced-order physical model, and a machine-learning residual corrector.
I also note that the definition of SoH must match the second-life application. A lithium ion battery used for low-power backup may tolerate capacity fade more than resistance rise. A lithium ion battery used for peak shaving may need high power and good thermal stability. Therefore, I recommend application-specific SoH definitions:
$$ SoH_{app} = g(C_{current}, R_{current}, \eta_{current}, T_{current}, \text{application profile}) $$
This application-aware definition prevents a single scalar SoH from hiding important limitations. It also supports multi-dimensional grading, where a retired lithium ion battery can be classified by capacity grade, power grade, and safety grade.
| Application | Dominant requirement | Preferred SoH indicator | Preferred estimation method |
|---|---|---|---|
| Telecommunication backup | Capacity and reliability | Capacity-based SoH | Partial capacity test plus model |
| Distributed microgrid | Cycle life and efficiency | Capacity and resistance | Hybrid ECM and data-driven |
| Low-speed electric vehicle | Power and energy | Resistance and capacity | Short pulse plus regression |
| Fast charging buffer | Power and thermal safety | Resistance and thermal health | EIS feature plus fuzzy logic |
| Stationary energy storage | Long-term capacity | Capacity fade trend | Mathematical model plus Bayesian update |
Conclusion
I conclude that SoH estimation for a retired lithium ion battery is moving from a single-dimension laboratory measurement toward a multi-modal, application-aware, and uncertainty-aware framework. Direct measurement methods remain the physical foundation. Model-based methods provide interpretability and online capability. Data-driven methods provide adaptability and nonlinear mapping. The most promising path is a data-physics synergy that uses fast measurements, reduced-order physical models, and machine learning to correct residuals. This synergy can reduce testing time, improve generalization, and provide the confidence needed for safe second-life deployment.
For future work, I recommend four priorities. First, develop multi-physics coupling models that are fast enough for online use. Second, integrate intelligent sensors that can capture local impedance, temperature, and strain in a lithium ion battery module. Third, design transferable and physics-informed algorithms that work with small labeled datasets. Fourth, standardize SoH definitions and uncertainty reporting so that retired lithium ion battery grading is comparable across suppliers and applications. If these priorities are addressed, retired lithium ion battery echelon utilization can become a scalable, safe, and economically viable part of the circular energy economy.