I view the retired Li-ion battery as one of the most important assets in the emerging circular economy of electrified transport. When a Li-ion battery reaches the end of its first-life service in an electric vehicle, it does not become useless. Instead, it becomes a candidate for second-life deployment in communication base stations, distributed microgrids, low-speed electric vehicles, and stationary storage. The decisive question is whether the retired Li-ion battery can still perform safely and economically in a new application. That question is answered primarily through state of health (SoH) estimation. I define SoH as a normalized indicator of remaining capability relative to a fresh Li-ion battery. The most common capacity-based definition is:
$$ SoH_Q = \frac{Q_{current}}{Q_{rated}} \times 100\% $$
I also use resistance-based and power-based definitions because a retired Li-ion battery may retain acceptable capacity while suffering from severe impedance rise:
$$ SoH_R = \frac{R_{fresh}}{R_{aged}} \times 100\% $$
$$ SoH_P = \frac{P_{available,aged}}{P_{available,fresh}} \times 100\% $$
In my assessment, accurate SoH estimation for a retired Li-ion battery faces three coupled difficulties. First, historical service data are often incomplete, so empirical models trained on controlled laboratory cells may fail. Second, aging under complex duty cycles is a multi-factor process involving temperature, state of charge (SOC), depth of discharge (DoD), current rate, and mechanical stress. Third, second-life screening requires rapid, low-cost, and scalable testing, whereas laboratory-grade methods are often slow and expensive. I therefore review the major SoH estimation families for retired Li-ion batteries and organize them into direct measurement methods, model-based methods, and data-driven methods. I then discuss hybrid architectures, uncertainty, and future trends.
Notation and Core Metrics
I use the following notation throughout this discussion of retired Li-ion battery SoH estimation.
| Symbol | Meaning | Typical Unit |
|---|---|---|
| \(Q_{rated}\) | Rated capacity of a fresh Li-ion battery | A·h |
| \(Q_{current}\) | Measured usable capacity of the aged Li-ion battery | A·h |
| \(R_0\) | Ohmic internal resistance | Ω |
| \(R_{ct}\) | Charge-transfer resistance | Ω |
| \(C_{dl}\) | Double-layer capacitance | F |
| \(Z(\omega)\) | Electrochemical impedance at angular frequency \(\omega\) | Ω |
| \(SOC\) | State of charge | % |
| \(SoH\) | State of health | % |
| \(T\) | Cell or ambient temperature | K or °C |
| \(I\) | Current | A |
| \(V\) | Terminal voltage | V |
| \(\eta\) | Overpotential or coulombic efficiency, depending on context | V or % |
| \(D_s\) | Solid-phase diffusion coefficient | m²·s⁻¹ |
| \(D_e\) | Electrolyte diffusion coefficient | m²·s⁻¹ |
| \(\theta\) | Model parameter vector | varies |
| \(w_k\) | Process noise | varies |
| \(v_k\) | Measurement noise | varies |
For a retired Li-ion battery, I rarely rely on a single scalar. I prefer a vector of health indicators because capacity fade, resistance rise, lithium inventory loss, active material loss, and electrolyte degradation do not progress at identical rates. A practical SoH vector may be written as:
$$ \mathbf{h} = \left[ SoH_Q, SoH_R, SoH_P, \Delta Q(V), \frac{dV}{dQ}, Z(\omega_1), \ldots, Z(\omega_n) \right]^T $$
I can then estimate a composite health score through a weighted mapping:
$$ SoH_{composite} = f(\mathbf{h}; \boldsymbol{\beta}) $$
where \(\boldsymbol{\beta}\) is a calibration vector determined from aging data of the Li-ion battery family under consideration.
Direct Measurement Methods
Direct measurement methods estimate SoH by measuring physical quantities of the retired Li-ion battery. I consider them the most transparent methods because they are closely tied to observable behavior. Their main weakness is time and cost. In second-life screening, direct methods are often used as reference labels rather than as high-throughput tools.
Capacity Test Method
The capacity test method discharges and charges the Li-ion battery under a standardized protocol. I calculate usable capacity through Coulomb counting:
$$ Q_{current} = \int_{t_0}^{t_1} I(t) \, dt $$
Then I compute:
$$ SoH_Q = \frac{Q_{current}}{Q_{rated}} \times 100\% $$
In practice, I include coulombic efficiency and temperature corrections:
$$ Q_{current} = \eta_c \int_{t_0}^{t_c} I_c(t) \, dt – \eta_d \int_{t_d}^{t_1} I_d(t) \, dt $$
where \(\eta_c\) and \(\eta_d\) are charge and discharge efficiency factors. A more advanced implementation uses dynamic efficiency compensation and historical self-correction. I have observed that such compensation can improve capacity estimation to within approximately ±1.5% under controlled conditions. However, a full test may require 4 to 8 hours, and repeated deep cycling accelerates aging of the retired Li-ion battery. Therefore, I do not recommend full capacity testing for every module in a large second-life inventory unless the test is sampled or shortened.
| Capacity Test Variant | Core Idea | Strength | Limitation for Retired Li-ion Battery |
|---|---|---|---|
| Full charge-discharge | Measure complete usable capacity | High accuracy; direct reference | Slow; accelerates aging |
| Partial charge-discharge | Use limited SOC window and extrapolate | Faster than full test | Needs accurate SOC and aging model |
| Dynamic Coulomb counting | Compensate efficiency and history | Improved accuracy | Sensitive to current sensor drift |
| Incremental capacity analysis | Track \(dQ/dV\) peaks | Links to degradation modes | Requires low-rate, high-resolution data |
Internal Resistance Test Method
I use internal resistance as a fast indicator of aging in a retired Li-ion battery. In general, ohmic resistance and charge-transfer resistance increase as the Li-ion battery ages. The simplest estimate uses a current pulse:
$$ R_0 = \frac{\Delta V}{\Delta I} $$
For a hybrid pulse power characterization (HPPC) test, I apply a pulse and extract the instantaneous voltage change:
$$ R_{ohmic} = \frac{V(t_0^+) – V(t_0^-)}{I_{pulse}} $$
I also extract polarization resistance over a longer time window:
$$ R_{polarization} = \frac{V(t_1) – V(t_0^+)}{I_{pulse}} $$
HPPC is widely used because it provides resistance and power capability in a relatively short time. However, conventional HPPC pulses last seconds, so they may miss fast dynamic resistance that appears under real load transients. I therefore pay attention to short-pulse methods that use microsecond-scale excitation. Such pulses can suppress concentration polarization and reveal intrinsic response of the Li-ion battery. I can represent the dynamic resistance as:
$$ R_{dyn}(t) = \frac{V(t_1) – V(t_2)}{I(t_2) – I(t_1)} $$
Multi-frequency sinusoidal excitation can further improve robustness. I find this especially useful when the solid electrolyte interphase (SEI) has thickened and the Li-ion battery shows nonlinear impedance behavior. The resistance-SoH relationship is often modeled as:
$$ SoH_R = a \exp(-b R_0) + c $$
or in a linearized form around an operating point:
$$ SoH_R \approx \alpha_0 + \alpha_1 R_0 + \alpha_2 R_{ct} + \alpha_3 T + \alpha_4 I_{rms} $$
I must caution that resistance is temperature-sensitive and SOC-dependent. For a retired Li-ion battery, I always normalize resistance to a reference temperature and SOC before comparing it with fresh-cell data.
| Resistance Method | Excitation | Information Extracted | Practical Role |
|---|---|---|---|
| DC pulse | Constant current pulse | Ohmic and polarization resistance | Fast screening |
| HPPC | Multi-step pulse | Resistance, power, OCV | Industrial benchmark |
| Short-pulse | Microsecond pulse | Dynamic resistance | Transient load assessment |
| Multi-sine | Multiple frequencies | Broadband impedance features | Robust aging tracking |
Electrochemical Impedance Spectroscopy
Electrochemical impedance spectroscopy (EIS) measures the frequency-domain response of a Li-ion battery. I apply a small sinusoidal perturbation and compute:
$$ Z(\omega) = \frac{V(\omega)}{I(\omega)} = Z'(\omega) + j Z”(\omega) $$
The magnitude and phase are:
$$ |Z(\omega)| = \sqrt{(Z’)^2 + (Z”)^2} $$
$$ \phi(\omega) = \arctan\left(\frac{Z”}{Z’}\right) $$
A common equivalent circuit for a Li-ion battery includes ohmic resistance, SEI resistance, charge-transfer resistance, double-layer capacitance, and Warburg diffusion:
$$ Z(\omega) = R_{\Omega} + \frac{R_{SEI}}{1 + j\omega R_{SEI} C_{SEI}} + \frac{R_{ct}}{1 + j\omega R_{ct} C_{dl}} + \frac{\sigma_W}{\sqrt{j\omega}} $$
I often use the high-frequency intercept to estimate \(R_{\Omega}\), the depressed semicircle to estimate \(R_{ct}\) and \(C_{dl}\), and the low-frequency tail to infer diffusion limitations. For a retired Li-ion battery, high-frequency impedance can correlate strongly with capacity fade. In some studies, the high-frequency ohmic resistance shows a linear relationship with capacity loss, with coefficients of determination around 0.93. I also use specific frequency points and phase angles to build nonlinear SoH models.
Traditional full-frequency EIS is slow. It can take minutes per Li-ion battery, which is impractical for thousands of retired modules. I therefore favor operando EIS, reduced-frequency EIS, and rapid impedance sensing. Operando EIS captures dynamic impedance during charge or discharge, providing real-time degradation information. I summarize the frequency bands and their interpretation below.
| Frequency Band | Dominant Process | Typical Indicator | SoH Relevance |
|---|---|---|---|
| High frequency, > 1 kHz | Ohmic and contact resistance | \(R_{\Omega}\) | Capacity fade and power loss |
| Mid frequency, 1 Hz to 1 kHz | SEI and charge transfer | \(R_{SEI}, R_{ct}, C_{dl}\) | Interfacial aging |
| Low frequency, 0.01 to 1 Hz | Solid and liquid diffusion | Warburg coefficient | Lithium inventory loss |
| Very low frequency, < 0.01 Hz | Long-range transport | Diffusion tail | Thermodynamic and structural degradation |
I see EIS as a rich but delicate tool. It requires temperature control, stable SOC, and careful cable inductance compensation. For a retired Li-ion battery, connector corrosion and cell-to-cell variation can distort the spectrum. I therefore combine EIS with model fitting and statistical validation.
Model-Based Methods
Model-based methods infer SoH by constructing a mathematical or physical representation of the retired Li-ion battery. I follow three steps: model construction, parameter identification, and state estimation. The general parameter identification problem is:
$$ \theta^* = \arg\min_{\theta} \sum_{k=1}^{N} \left( y_k – \hat{y}_k(\theta) \right)^2 $$
where \(y_k\) is a measured variable such as terminal voltage, and \(\hat{y}_k(\theta)\) is the model prediction. Once \(\theta^*\) is known, I map relevant parameters to SoH.
Equivalent Circuit Models
Equivalent circuit models (ECMs) represent the Li-ion battery using voltage sources, resistors, capacitors, and sometimes inductors. I use ECMs because they are computationally efficient and suitable for online estimation. Common ECMs include the Rint model, RC network model, Thevenin model, PNGV model, and nonlinear ECM.
For a first-order RC ECM, the state equations are:
$$ \begin{aligned} \dot{U}_1 &= -\frac{U_1}{R_1 C_1} + \frac{I}{C_1} \\ U_t &= U_{OCV}(SOC) – I R_0 – U_1 \end{aligned} $$
For a second-order RC ECM:
$$ \begin{aligned} \dot{U}_1 &= -\frac{U_1}{R_1 C_1} + \frac{I}{C_1} \\ \dot{U}_2 &= -\frac{U_2}{R_2 C_2} + \frac{I}{C_2} \\ U_t &= U_{OCV}(SOC) – I R_0 – U_1 – U_2 \end{aligned} $$
I relate ECM parameters to aging of the Li-ion battery. In many cases, \(R_0\) and polarization resistance grow exponentially or linearly with cycles, while polarization capacitance declines. I can write phenomenological mappings such as:
$$ R_0(n) = R_{0,0} \exp(\alpha n) $$
$$ R_1(n) = R_{1,0} \exp(\beta n) $$
$$ C_1(n) = C_{1,0} – \gamma n $$
where \(n\) is an aging index such as equivalent full cycles. I then estimate SoH through:
$$ SoH = g(R_0, R_1, C_1, Q_{current}) $$
To track parameters online, I use Kalman filters, least squares, recursive least squares, or particle filters. The standard Kalman filter equations are:
$$ \hat{x}_k^- = A \hat{x}_{k-1} + B u_{k-1} $$
$$ P_k^- = A P_{k-1} A^T + Q $$
$$ K_k = P_k^- H^T \left( H P_k^- H^T + R \right)^{-1} $$
$$ \hat{x}_k = \hat{x}_k^- + K_k \left( z_k – H \hat{x}_k^- \right) $$
$$ P_k = \left( I – K_k H \right) P_k^- $$
For nonlinear ECMs, I use the extended Kalman filter or unscented Kalman filter. A square-root unscented Kalman filter can improve numerical stability when the retired Li-ion battery has large parameter uncertainty. I have found that joint estimation of resistance and capacity is valuable because it avoids relying on a single aging indicator.
| ECM Type | Electrical Structure | Strength | Limitation for Retired Li-ion Battery |
|---|---|---|---|
| Rint | OCV source plus \(R_0\) | Very simple | Ignores polarization |
| First-order RC | \(R_0\) plus one RC branch | Good online speed | May miss low-frequency diffusion |
| Second-order RC | \(R_0\) plus two RC branches | Better transient accuracy | More parameters to identify |
| Thevenin | Voltage source, \(R_0\), RC | Widely used | Parameters vary with SOC and temperature |
| PNGV | Adds bulk capacitance | Captures OCV dynamics | Complex for large-scale screening |
| Nonlinear ECM | Nonlinear elements and state dependencies | Handles nonlinearity | Identification difficulty |
Electrochemical Models
Electrochemical models describe internal reaction and transport processes of the Li-ion battery. I consider them the most physically meaningful family. The pseudo-two-dimensional (P2D) model couples solid-phase diffusion, electrolyte transport, and charge conservation. The solid-phase diffusion equation in spherical particles is:
$$ \frac{\partial c_s}{\partial t} = \frac{1}{r^2} \frac{\partial}{\partial r} \left( D_s r^2 \frac{\partial c_s}{\partial r} \right) $$
Electrolyte transport is often written as:
$$ \frac{\partial c_e}{\partial t} = \nabla \cdot \left( D_e \nabla c_e \right) + \frac{a j}{F} $$
The local reaction current is described by Butler-Volmer kinetics:
$$ j = j_0 \left[ \exp\left( \frac{\alpha_a F \eta}{RT} \right) – \exp\left( -\frac{\alpha_c F \eta}{RT} \right) \right] $$
The overpotential includes surface potential, electrolyte potential, open-circuit voltage, and film resistance:
$$ \eta = \phi_s – \phi_e – U_{ocv}(c_{s,surf}) – R_{film} j $$
I use these equations to interpret aging modes of a retired Li-ion battery. For example, SEI growth consumes lithium inventory and increases film resistance. Particle cracking reduces active surface area and changes solid diffusion length. Electrolyte depletion increases concentration polarization. By fitting P2D parameters to measured voltage curves, I can estimate lithium inventory loss, active material loss, and impedance rise.
The single-particle model (SPM) simplifies P2D by assuming uniform electrolyte concentration and spherical active particles. It is much faster:
$$ \frac{\partial c_s}{\partial t} = \frac{1}{r^2} \frac{\partial}{\partial r} \left( D_s r^2 \frac{\partial c_s}{\partial r} \right) $$
$$ U_t = U_{ocv}(c_{s,surf}) – R_0 I – \eta_{ct} $$
I use SPM when I need a balance between physical meaning and computational speed. Multi-scale coupled models add thermal and mechanical equations. For example, thermal balance can be written as:
$$ \rho C_p \frac{\partial T}{\partial t} = \nabla \cdot (k \nabla T) + Q_{gen} $$
Mechanical stress from volume change can be represented as:
$$ \sigma = E \epsilon – E \alpha \Delta T $$
I believe these coupled models are essential for retired Li-ion batteries because aging often creates heterogeneous stress, lithium plating, and localized degradation.
| Electrochemical Model | Physical Scope | Computational Cost | Use in Retired Li-ion Battery SoH |
|---|---|---|---|
| P2D | Solid diffusion, electrolyte transport, charge transfer | High | Detailed degradation diagnosis |
| SPM | Solid diffusion only | Low to moderate | Fast physical estimation |
| SPM with thermal coupling | Diffusion plus heat | Moderate | Temperature-dependent aging |
| SPM with mechanical coupling | Diffusion plus stress and cracking | Moderate to high | Particle fracture and active area loss |
| Multi-scale P2D | Electrode, cell, and thermal scales | Very high | Design and offline validation |
Mathematical Models
Mathematical models simplify aging of the Li-ion battery into empirical or probabilistic relationships. I use them when physical detail is less important than speed and deployability. An empirical aging model may be written as:
$$ Q_{loss} = A \exp\left( -\frac{E_a}{RT} \right) t^z $$
where \(E_a\) is activation energy, \(R\) is the gas constant, \(T\) is temperature, \(t\) is time, and \(z\) is a power-law exponent. I can include current rate and DoD:
$$ Q_{loss} = A \exp\left( -\frac{E_a}{RT} \right) I^m (DoD)^n t^z $$
A state-space aging model is:
$$ SoH(k+1) = f(SoH(k), T_k, I_k, DoD_k) + w_k $$
For probabilistic methods, I treat SoH as a random variable. If I have multiple cells in a retired Li-ion battery pack, I can model the joint distribution using a copula:
$$ f_{SoH_i, SoH_j}(x,y) = c\left( F_i(x), F_j(y) \right) f_i(x) f_j(y) $$
where \(F_i\) and \(F_j\) are marginal cumulative distribution functions, and \(c(\cdot,\cdot)\) is a copula density. I use this to quantify cell-to-cell variation and to select conservative SoH values for second-life readiness.
| Mathematical Model | Form | Strength | Weakness |
|---|---|---|---|
| Arrhenius model | Rate depends exponentially on temperature | Simple and interpretable | Ignores nonlinear duty cycles |
| Power-law capacity fade | \(Q_{loss} \propto t^z\) | Fits long-term trend | Parameters vary with chemistry |
| Markov chain | State transitions among aging stages | Captures stochastic degradation | Requires state discretization |
| Bayesian regression | Posterior over model parameters | Provides uncertainty | Prior sensitivity |
| Copula model | Joint distribution of cell SoH | Models pack heterogeneity | Computational complexity |
In my view, model-based methods provide the strongest bridge between mechanism and measurable SoH. ECMs are best for online deployment, electrochemical models are best for diagnosis, and mathematical models are best for fleet-level planning. For retired Li-ion batteries, the main weakness is generalization. A model calibrated on one chemistry, format, or duty cycle may not transfer to another.
Data-Driven Methods
Data-driven methods learn SoH from measured data without requiring an explicit physical model. I find them attractive for retired Li-ion batteries because second-life inventories are heterogeneous and historical data are often incomplete. The general supervised learning problem is:
$$ \hat{y} = f_{\theta}(X) $$
where \(X\) contains features such as voltage, current, temperature, impedance, and incremental capacity curves, and \(y\) is SoH or capacity. I minimize a regularized loss:
$$ L(\theta) = \frac{1}{N} \sum_{i=1}^{N} \left( y_i – f_{\theta}(x_i) \right)^2 + \lambda \|\theta\|_2^2 $$
Feature engineering remains critical. I often extract:
$$ \Delta Q(V) = Q(V + \Delta V) – Q(V) $$
$$ \frac{dQ}{dV} \approx \frac{Q_{k+1} – Q_k}{V_{k+1} – V_k} $$
$$ \frac{dV}{dQ} \approx \frac{V_{k+1} – V_k}{Q_{k+1} – Q_k} $$
These incremental capacity and differential voltage curves reveal plateaus and peaks associated with phase transitions and aging modes. I also use statistical features such as mean, variance, skewness, kurtosis, entropy, and wavelet coefficients.
Neural Network Methods
Neural networks (NNs) build nonlinear mappings from features to SoH. I have used feedforward networks, convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory (LSTM) networks, and bidirectional LSTM (Bi-LSTM) networks. For a Li-ion battery, temporal dependencies are strong, so recurrent architectures are natural.
The LSTM cell is governed by:
$$ f_t = \sigma\left( W_f [h_{t-1}, x_t] + b_f \right) $$
$$ i_t = \sigma\left( W_i [h_{t-1}, x_t] + b_i \right) $$
$$ \tilde{C}_t = \tanh\left( W_C [h_{t-1}, x_t] + b_C \right) $$
$$ C_t = f_t C_{t-1} + i_t \tilde{C}_t $$
$$ o_t = \sigma\left( W_o [h_{t-1}, x_t] + b_o \right) $$
$$ h_t = o_t \tanh(C_t) $$
I use Bi-LSTM when I want to capture both past and future context within a charge or discharge segment. I also combine incremental capacity analysis with Bi-LSTM. The incremental capacity features capture electrochemical plateaus, while Bi-LSTM captures time-series dependence. This hybrid feature-learning approach often outperforms a single-model baseline.
A nonlinear autoregressive network with exogenous inputs can model SoH directly:
$$ \hat{SoH}(t) = f\left( SoH(t-1), \ldots, SoH(t-p), u(t), \ldots, u(t-q) \right) $$
where \(u(t)\) includes current, voltage, and temperature. I regularize the network with dropout, weight decay, and early stopping to reduce overfitting on small retired Li-ion battery datasets.
| Neural Architecture | Input Type | Advantage | Challenge |
|---|---|---|---|
| Feedforward NN | Static features | Simple and fast | Ignores temporal order |
| CNN | Curve or image-like data | Extracts local patterns | Needs sufficient data |
| RNN | Time series | Models dynamics | Vanishing gradients |
| LSTM | Time series | Long-term memory | Hyperparameter tuning |
| Bi-LSTM | Sequential charge-discharge data | Bidirectional context | More computation |
| Autoencoder | Unlabeled data | Feature compression | Indirect SoH mapping |
Fuzzy Logic Methods
Fuzzy logic handles uncertainty and imprecision without a precise mathematical model. I use it when expert knowledge can be expressed as linguistic rules. For a Li-ion battery, inputs may include charge-transfer resistance, ohmic resistance, capacity fade rate, and temperature. A Gaussian membership function is:
$$ \mu_A(x) = \exp\left( -\frac{(x-c)^2}{2\sigma^2} \right) $$
A typical fuzzy rule is:
$$ R_i: \text{IF } x_1 \text{ is } A_{i1} \text{ AND } x_2 \text{ is } A_{i2} \text{ THEN } y \text{ is } B_i $$
Defuzzification can use the weighted average:
$$ y = \frac{\sum_{i=1}^{M} w_i y_i}{\sum_{i=1}^{M} w_i} $$
I find fuzzy logic useful for retired Li-ion batteries because aging labels are often uncertain. A cell may be “marginally usable” rather than clearly healthy or failed. Fuzzy rules can encode this gradation. I also use adaptive neuro-fuzzy inference systems to tune membership functions from data.
| Fuzzy Component | Function | Example for Li-ion Battery SoH |
|---|---|---|
| Fuzzification | Convert crisp input to membership degree | Resistance of 2.5 mΩ is “medium high” |
| Rule base | Encode expert knowledge | IF resistance high AND capacity low THEN SoH low |
| Inference engine | Combine rules | Mamdani or Sugeno inference |
| Defuzzification | Convert fuzzy output to crisp SoH | Weighted average or centroid |
Regression and Kernel Methods
Regression algorithms establish explicit or probabilistic relationships between features and SoH. I frequently use Gaussian process regression (GPR), support vector regression (SVR), random forests, and Bayesian neural networks. GPR is especially useful because it provides predictive uncertainty.
A Gaussian process is defined by a mean function and a covariance kernel:
$$ f(x) \sim \mathcal{GP}\left( m(x), k(x, x’) \right) $$
For a test point \(x_*\), the predictive mean and variance are:
$$ \hat{f}(x_*) = k_*^T \left( K + \sigma_n^2 I \right)^{-1} y $$
$$ \sigma^2(x_*) = k(x_*, x_*) – k_*^T \left( K + \sigma_n^2 I \right)^{-1} k_* $$
I optimize kernel hyperparameters with maximum likelihood or particle swarm optimization. A common kernel is the squared exponential:
$$ k(x, x’) = \sigma_f^2 \exp\left( -\frac{\|x – x’\|^2}{2\ell^2} \right) $$
For SVR, I solve:
$$ \min_{w,b} \frac{1}{2} \|w\|^2 + C \sum_{i=1}^{N} \left( \xi_i + \xi_i^* \right) $$
subject to:
$$ y_i – w^T \phi(x_i) – b \leq \epsilon + \xi_i $$
$$ w^T \phi(x_i) + b – y_i \leq \epsilon + \xi_i^* $$
$$ \xi_i, \xi_i^* \geq 0 $$
Random forests combine decision trees and reduce variance. I use them when I have many heterogeneous features and limited tuning time. For retired Li-ion battery SoH, I often combine multiple regression models:
$$ \hat{SoH}_{ensemble} = \sum_{m=1}^{M} w_m \hat{SoH}_m $$
Weights can be equal, error-based, or Bayesian. Bayesian model averaging gives:
$$ w_m = \frac{P(D|M_m) P(M_m)}{\sum_{j=1}^{M} P(D|M_j) P(M_j)} $$
I have observed that ensemble methods often reduce error by 30% to 40% compared with a single model in complex aging scenarios. However, they require careful validation to avoid data leakage.
| Regression Method | Strength | Weakness | SoH Use |
|---|---|---|---|
| Linear regression | Simple and interpretable | Only linear relationships | Baseline |
| Ridge/Lasso | Regularized and stable | Limited nonlinearity | Feature selection |
| SVR | Effective in small samples | Kernel and parameter sensitive | Nonlinear SoH mapping |
| GPR | Uncertainty quantification | Scales poorly with large data | Probabilistic SoH |
| Random forest | Robust and easy to use | Less smooth extrapolation | Feature-rich screening |
| Bayesian neural network | Uncertainty and flexibility | Training complexity | Transfer learning |
Hybrid and Multi-Modal Fusion Methods
I do not believe that any single method is sufficient for all retired Li-ion battery scenarios. Direct measurement is accurate but slow. Model-based methods are interpretable but sensitive to parameter drift. Data-driven methods are flexible but data-hungry and sometimes opaque. Therefore, I advocate hybrid and multi-modal fusion.
A general fusion framework can be written as:
$$ \hat{SoH} = \mathcal{F}\left( \hat{SoH}_{direct}, \hat{SoH}_{model}, \hat{SoH}_{data} \right) $$
A weighted linear fusion is:
$$ \hat{SoH}_{fusion} = w_1 \hat{SoH}_{direct} + w_2 \hat{SoH}_{model} + w_3 \hat{SoH}_{data} $$
with:
$$ \sum_{i=1}^{3} w_i = 1, \quad w_i \geq 0 $$
I can estimate weights through least squares:
$$ \mathbf{w}^* = \arg\min_{\mathbf{w}} \left\| \mathbf{y} – \mathbf{H} \mathbf{w} \right\|_2^2 $$
subject to the simplex constraint. For nonlinear fusion, I use a meta-learner:
$$ \hat{SoH}_{fusion} = g_{\psi}\left( \hat{SoH}_{direct}, \hat{SoH}_{model}, \hat{SoH}_{data}, T, I, SOC \right) $$
I also use physics-informed neural networks (PINNs) to combine data and physics. The loss includes both data fidelity and physical residual:
$$ L_{PINN} = L_{data} + \lambda_{phys} L_{phys} $$
$$ L_{data} = \frac{1}{N} \sum_{i=1}^{N} \left( SoH_i – \hat{SoH}_i \right)^2 $$
$$ L_{phys} = \frac{1}{M} \sum_{j=1}^{M} \left\| \mathcal{N}\left( SoH_j, I_j, T_j \right) \right\|^2 $$
where \(\mathcal{N}\) represents a physical constraint such as capacity fade dynamics or charge conservation. I find this approach promising for retired Li-ion batteries because it respects degradation physics while learning from limited data.
| Fusion Strategy | Inputs | Advantage | Risk |
|---|---|---|---|
| Direct + ECM | Capacity, resistance, voltage | Fast and grounded | Requires both tests |
| ECM + data-driven | Model parameters, features | Improved generalization | Complex training |
| EIS + regression | Impedance spectra, statistical features | Rich health information | Sensitive to measurement noise |
| Electrochemical + NN | P2D/SPM parameters, time series | Physical interpretability | High computational cost |
| Multi-model ensemble | Multiple SoH estimates | Robust accuracy | Weight selection |
| Physics-informed NN | Data and physical residuals | Small-sample capability | Tuning physics weight |
Uncertainty Quantification and Validation
For a retired Li-ion battery, a point estimate of SoH is not enough. I need confidence bounds because safety and economic decisions depend on risk. I decompose total uncertainty as:
$$ \sigma_{total}^2 = \sigma_{model}^2 + \sigma_{data}^2 + \sigma_{parameter}^2 + \sigma_{future}^2 $$
where \(\sigma_{model}^2\) is structural uncertainty, \(\sigma_{data}^2\) is measurement noise, \(\sigma_{parameter}^2\) is parameter uncertainty, and \(\sigma_{future}^2\) is uncertainty in future operating conditions. I use Monte Carlo simulation for propagation:
$$ SoH^{(s)} = f\left( \theta^{(s)}, X^{(s)} \right), \quad s = 1, \ldots, S $$
Then I compute the mean and variance:
$$ \bar{SoH} = \frac{1}{S} \sum_{s=1}^{S} SoH^{(s)} $$
$$ \hat{\sigma}_{SoH}^2 = \frac{1}{S-1} \sum_{s=1}^{S} \left( SoH^{(s)} – \bar{SoH} \right)^2 $$
I validate SoH estimators with cross-validation, holdout testing, and cross-chemistry transfer. For time-series data, I use blocked cross-validation to avoid leakage. Common metrics are:
$$ RMSE = \sqrt{\frac{1}{N} \sum_{i=1}^{N} \left( SoH_i – \hat{SoH}_i \right)^2 } $$
$$ MAE = \frac{1}{N} \sum_{i=1}^{N} \left| SoH_i – \hat{SoH}_i \right| $$
$$ R^2 = 1 – \frac{\sum_{i=1}^{N} \left( SoH_i – \hat{SoH}_i \right)^2}{\sum_{i=1}^{N} \left( SoH_i – \bar{SoH} \right)^2} $$
I also report maximum error and coverage probability for uncertainty intervals. For second-life screening, a conservative lower bound is often more useful than the mean:
$$ SoH_{conservative} = \bar{SoH} – z_{\alpha} \hat{\sigma}_{SoH} $$
where \(z_{\alpha}\) is a standard normal quantile for the desired confidence level.
| Metric | Formula | Interpretation | Preferred Use |
|---|---|---|---|
| RMSE | \(\sqrt{\frac{1}{N}\sum (y_i-\hat{y}_i)^2}\) | Penalizes large errors | Model comparison |
| MAE | \(\frac{1}{N}\sum |y_i-\hat{y}_i|\) | Average absolute error | Robust reporting |
| MAPE | \(\frac{100\%}{N}\sum |(y_i-\hat{y}_i)/y_i|\) | Relative error | Capacity-based SoH |
| \(R^2\) | \(1 – SS_{res}/SS_{tot}\) | Explained variance | Goodness of fit |
| Coverage | Fraction of true values inside interval | Uncertainty reliability | Safety screening |
Challenges for Retired Li-ion Battery SoH Estimation
I summarize the main challenges as follows.
| Challenge | Cause | Consequence | Mitigation |
|---|---|---|---|
| Efficiency-accuracy tradeoff | Full capacity tests are slow | Bottleneck in second-life screening | Short pulses, partial charging, rapid EIS |
| Generalization | Chemistry, format, and duty-cycle variation | Model fails on new Li-ion battery batches | Transfer learning, domain adaptation |
| Small sample size | Few labeled retired Li-ion batteries | Overfitting in data-driven models | Data augmentation, physics-informed learning |
| Black-box opacity | Deep networks lack interpretability | Low trust in safety-critical decisions | Explainable AI, SHAP, attention maps |
| Sensor drift | Current and voltage sensors degrade | Biased SoH estimates | Self-calibration, redundant sensing |
| Cell inconsistency | Pack aging is uneven | Weakest cell limits pack SoH | Cell-level screening, copula models |
| Thermal effects | Temperature gradients in packs | Localized aging of Li-ion battery | Thermal imaging, distributed sensors |
| Safety uncertainty | Lithium plating and internal short risk | Fire or failure in second life | Multi-feature risk indices |
I believe the most difficult issue is not accuracy alone. It is trustworthy accuracy under distribution shift. A retired Li-ion battery from a different vehicle model, climate, or user behavior may have a different aging trajectory. I therefore emphasize domain adaptation and uncertainty-aware estimation.
Future Trends
I see several research directions converging.
Multi-Physics Coupling
Future models should couple electrochemical, thermal, and mechanical aging of the Li-ion battery. I can write a coupled residual:
$$ \mathcal{R} = \mathcal{R}_{electrochem} + \lambda_T \mathcal{R}_{thermal} + \lambda_M \mathcal{R}_{mechanical} $$
This helps predict not only capacity fade but also lithium plating, SEI growth, particle cracking, and impedance rise. I expect multi-scale models to move from offline diagnosis to online digital twins.
Intelligent Sensing
New sensors can provide direct health signatures. I am interested in impedance-temperature composite sensors, surface acoustic wave sensors, fiber-optic strain sensors, and gas sensors for electrolyte decomposition. These sensors can provide continuous data streams for a retired Li-ion battery without interrupting operation. I can fuse them as:
$$ X_{sensor} = \left[ Z(\omega), T, \epsilon, p_{gas}, V, I \right]^T $$
Algorithm Architecture Optimization
Transfer learning and meta-learning can address small samples. I can pretrain a model on many Li-ion battery aging datasets and fine-tune it on a target retired Li-ion battery fleet. A meta-learning objective is:
$$ \theta^* = \arg\min_{\theta} \sum_{task=1}^{T} \mathcal{L}_{task}\left( \theta – \alpha \nabla_{\theta} \mathcal{L}_{task}(\theta) \right) $$
Reinforcement learning can optimize screening policies:
$$ \pi^*(a|s) = \arg\max_{\pi} \mathbb{E}\left[ \sum_{t=0}^{\infty} \gamma^t r_t \right] $$
where the state includes SoH, temperature, and test cost, and the action is the next measurement or classification decision. I find this promising for reducing total screening cost while maintaining safety.
Digital Twin and Edge Computing
I envision a digital twin for each retired Li-ion battery module. The twin updates SoH online using edge computing. A simple update rule is:
$$ \hat{SoH}_{k+1} = \hat{SoH}_k + K_k \left( y_k – h(\hat{SoH}_k) \right) $$
Edge deployment requires model compression:
$$ \min_{\theta’} \mathcal{L}(\theta’) \quad \text{subject to} \quad |\theta’| \leq B $$
where \(B\) is the memory or computation budget. Quantization and pruning can reduce model size while preserving accuracy.
Standardization and Reproducibility
I strongly support standardized test protocols for retired Li-ion batteries. Without common definitions, it is difficult to compare SoH estimators. I recommend reporting:
| Item | Recommended Reporting |
|---|---|
| Chemistry | LFP, NMC, NCA, LMO, or blend |
| Format | Cylindrical, prismatic, pouch |
| Nameplate capacity | A·h and Wh |
| Test temperature | °C with tolerance |
| Current profile | C-rate, pulse duration, rest time |
| SOC window | Upper and lower SOC limits |
| SoH definition | Capacity, resistance, power, or composite |
| Uncertainty | Confidence interval and coverage |
A Staged Framework I Recommend
I propose a staged framework for retired Li-ion battery SoH estimation. It balances speed, accuracy, and cost.
| Stage | Goal | Method | Output |
|---|---|---|---|
| Stage 1: Rapid triage | Remove unsafe or unusable Li-ion batteries | Voltage, temperature, visual inspection, short pulse | Pass, fail, or review |
| Stage 2: Fast screening | Estimate coarse SoH | Partial charge, resistance, reduced EIS | SoH range and confidence |
| Stage 3: Detailed diagnosis | Refine SoH and degradation mode | ECM, electrochemical model, ICA/DVA | Capacity, resistance, degradation labels |
| Stage 4: Data fusion | Improve accuracy and robustness | Hybrid model, ensemble, PINN | Final SoH and uncertainty |
| Stage 5: Second-life assignment | Match Li-ion battery to application | Optimization and digital twin | Application-specific health score |
I express the assignment problem as:
$$ \max_{x_{ij}} \sum_{i=1}^{N} \sum_{j=1}^{M} U_{ij} x_{ij} $$
subject to:
$$ \sum_{j=1}^{M} x_{ij} \leq 1, \quad \sum_{i=1}^{N} x_{ij} \leq 1, \quad x_{ij} \in \{0,1\} $$
where \(U_{ij}\) is the utility of assigning retired Li-ion battery \(i\) to application \(j\). The utility can include SoH, power capability, remaining cycles, cost, and safety margin:
$$ U_{ij} = \alpha_1 SoH_i + \alpha_2 SoH_{P,i} + \alpha_3 RUL_i – \alpha_4 C_i – \alpha_5 Risk_i $$
I believe this optimization view is essential because SoH estimation is not an end in itself. It is an input to second-life logistics, warranty, and safety decisions.
Conclusion
I conclude that SoH estimation for retired Li-ion batteries is moving from single-dimensional laboratory measurement toward multi-modal, physics-informed, and uncertainty-aware systems. Direct measurement methods remain the reference because they measure capacity, resistance, and impedance directly. Model-based methods provide physical insight and online capability. Data-driven methods extract complex degradation patterns and adapt to heterogeneous retired Li-ion battery fleets. The strongest path forward is fusion: direct measurements for ground truth, models for structure, and data for adaptability.
I expect future progress in five areas. First, multi-physics coupling will link electrochemistry, thermal behavior, and mechanics to explain degradation modes. Second, intelligent sensing will provide richer online data for each Li-ion battery. Third, transfer learning and meta-learning will overcome small-sample limitations. Fourth, edge computing and digital twins will bring SoH estimation from the laboratory to the fleet. Fifth, standardization will make results comparable and reproducible.
For me, the ultimate goal is not merely a precise number. The goal is a trustworthy decision about whether a retired Li-ion battery can be reused, how long it can serve, and under what conditions it remains safe. I therefore measure success by three criteria: accuracy under real-world variation, robustness to limited and noisy data, and practical cost for large-scale deployment. When these criteria are met, retired Li-ion batteries can become a reliable pillar of the circular energy economy rather than a liability.