The rapid global adoption of electric vehicles has driven an exponential increase in the manufacturing and deployment of lithium-ion batteries. Once an EV battery reaches the end of its first automotive service life, it may still retain a significant portion of its original capacity and thus can be repurposed in less demanding applications such as stationary energy storage, low-speed electric vehicles, and backup power systems. This cascade utilization of retired EV battery packs is a key strategy for improving resource efficiency and reducing the environmental burden of battery waste. However, the integration of retired cells into second-life applications introduces distinct safety challenges. The internal state of an aged EV battery is far more complex than that of a fresh cell: the electrode structure has evolved, lithium inventory has been partially consumed, and heterogeneous side reactions have altered the interface chemistry. Accurately diagnosing failure behaviors in such cells is therefore essential.

In this dissertation, I focus on second-life lithium iron phosphate batteries and develop a systematic framework for on-line diagnosis of failure behavior. The framework is built upon a failure physical model that couples the main intercalation/deintercalation reactions with parasitic side reactions such as solid electrolyte interphase growth and lithium plating. I evaluate off-line and on-line parameter identification approaches, explore the aging characteristics under different load profiles, construct dynamic failure boundaries based on parameter evolution, and validate the diagnostic results with multi-scale post-mortem analysis. My goal is to bridge the gap between physically meaningful model parameters and actionable safety state indicators for second-life EV battery management.
1. Development of a Failure Physical Model for the EV Battery
To describe the internal behavior of an operating lithium-ion cell, I employed a simplified electrochemical model based on the single-particle-plus framework. The model treats the positive and negative electrodes as two representative spherical particles and includes solid-phase diffusion, liquid-phase concentration gradients, reaction polarization, concentration polarization, and ohmic drop. The model equations are summarized below.
| Process | Equation |
|---|---|
| State of charge | $$\mathrm{SOC}(t)=1-\frac{I(t)t}{Q_{\mathrm{all}}}$$ |
| Average lithium fraction | $$y_{\mathrm{avg}} = y_0 + (1-\mathrm{SOC})\frac{Q_{\mathrm{all}}}{Q_p}, \quad x_{\mathrm{avg}} = x_0 – (1-\mathrm{SOC})\frac{Q_{\mathrm{all}}}{Q_n}$$ |
| Surface concentration deviation | $$\Delta y(t) = \Delta y(t_k) + \frac{I(t)\tau_p}{Q_p/7} + \frac{12\Delta y(t_k)\tau_p}{t-t_k}, \quad \Delta x(t) = \Delta x(t_k) – \frac{I(t)\tau_n}{Q_n/7} – \frac{12\Delta x(t_k)\tau_n}{t-t_k}$$ |
| Concentration polarization overpotential | $$\eta_{\mathrm{con}}(t) = \frac{2RT}{F}\left(1-t_+\right)\ln\frac{c_0+\Delta c(t)}{c_0-\Delta c(t)}$$ |
| Activation polarization overpotential | $$\eta_{\mathrm{act}}(t) = \frac{2RT}{F}\left[\ln\left(m_n+\sqrt{1+m_n^2}\right)+\ln\left(m_p+\sqrt{1+m_p^2}\right)\right]$$ |
| Ohmic overpotential | $$\eta_{\mathrm{ohm}}(t) = R_{\mathrm{ohm}} I(t)$$ |
| Terminal voltage | $$U(t) = U_p\big(y_{\mathrm{surf}}(t)\big) – U_n\big(x_{\mathrm{surf}}(t)\big) – \eta_{\mathrm{con}}(t) – \eta_{\mathrm{act}}(t) – \eta_{\mathrm{ohm}}(t)$$ |
The key parameters in the above model include the initial stoichiometric windows \(y_0\) and \(x_0\), the individual electrode capacities \(Q_p\) and \(Q_n\), the total cell capacity \(Q_{\mathrm{all}}\), the solid phase diffusion time constants \(\tau_p\) and \(\tau_n\), the liquid phase diffusion time constant \(\tau_e\), the diffusion proportional coefficient \(P_{\mathrm{con}}\), the reaction polarization coefficient \(P_{\mathrm{act}}\), and the ohmic resistance \(R_{\mathrm{ohm}}\). These parameters are not fixed during the lifetime of an EV battery: electrode structural degradation, interfacial film growth, and active material loss continuously modify them. Therefore, they can serve as indirect indicators of internal failure mechanisms.
1.1 SEI Film Growth Sub-Model
The solid electrolyte interphase film grows on the graphite negative electrode as the electrolyte is reduced. It consumes active lithium and increases internal resistance. I coupled the SEI side-reaction current density \(j_s\) with the main intercalation current density \(j_f\), where the total current density satisfies \(j = j_f + j_s\). The main reaction follows the Butler-Volmer equation while the SEI formation is described by an asymmetric Tafel-like expression because of its irreversible character:
$$j_s(t) = -a_s i_{\mathrm{SEI}}\exp\!\left(-\frac{\alpha_s F}{RT}\eta_{\mathrm{SEI}}(t)\right),$$
with the SEI overpotential
$$\eta_{\mathrm{SEI}}(t) = \eta_m(t) + U_n(t) – U_{n,\mathrm{ref}}(t) – \frac{R_{\mathrm{SEI}}(t)}{a_s} j_s(t).$$
The film thickness \(\delta_{\mathrm{SEI}}\), resistance \(R_{\mathrm{SEI}}\), and total lithium inventory \(Q_{\mathrm{Li}}\) evolution are described by:
$$\frac{\partial \delta_{\mathrm{SEI}}}{\partial t} = – \frac{M_{\mathrm{SEI}}}{\rho_{\mathrm{SEI}} a_s F} j_s, \quad R_{\mathrm{SEI}} = R_{\mathrm{SEI}}^0 + \frac{\delta_{\mathrm{SEI}}}{\kappa_{\mathrm{SEI}}},$$
$$\frac{d Q_{\mathrm{Li}}}{dt} = A \delta_n j_s, \quad Q_{\mathrm{Li}} = Q_p y_0 + Q_n x_0.$$
The SEI side reaction consumes active lithium, so \(Q_{\mathrm{Li}}\) decreases over time. The parameters used in the SEI model are summarized below.
| Parameter | Symbol | Value |
|---|---|---|
| Electrode area | \(A\) | 0.1 m² |
| Specific surface area of anode | \(a_s\) | \(1.8 \times 10^5\) m⁻¹ |
| Anode thickness | \(\delta_n\) | \(1 \times 10^{-4}\) m |
| Transfer coefficient | \(\alpha_s\) | 0.5 |
| Molar mass of SEI | \(M_{\mathrm{SEI}}\) | 0.162 kg mol⁻¹ |
| Density of SEI | \(\rho_{\mathrm{SEI}}\) | 1690 kg m⁻³ |
| Ionic conductivity of SEI | \(\kappa_{\mathrm{SEI}}\) | \(5\times 10^{-6}\) m s⁻¹ |
| Reference potential of side reaction | \(U_{\mathrm{ref},s}\) | 0.4 V |
| Initial exchange current factor | \(k\) | \(2\times10^{-6}\) A m²⁵ mol⁻¹⁵ |
| Faraday constant | \(F\) | \(9.6487\times10^{4}\) C mol⁻¹ |
1.2 Lithium Plating Sub-Model
At low temperature or high charge rates, lithium plating may occur on the graphite surface. Plating is especially hazardous because dendrites can pierce the separator and lead to internal short circuits. I account for lithium plating with a current density \(j_{\mathrm{Li}}\) described by a Butler-Volmer expression modified with a minimum function to ensure that lithium deposition only occurs when the thermodynamic condition is favorable:
$$j_{\mathrm{Li}}(t) = -a_s i_{\mathrm{Li}} \min\left\{0,\, \exp\!\left(\frac{\alpha_a F}{RT}\eta_{\mathrm{Li}}\right) – \exp\!\left(-\frac{\alpha_c F}{RT}\eta_{\mathrm{Li}}\right)\right\},$$
where
$$\eta_{\mathrm{Li}}(t) = \eta_m(t) + U_n(t) – U_{\mathrm{ref},s},\quad U_{\mathrm{ref},s}=0.$$
Similar to the SEI film, the plated lithium layer has a resistance \(R_{\mathrm{Li}}\), and the active lithium loss due to plating is
$$\frac{dQ_{\mathrm{Li}}}{dt} = A\delta_n j_{\mathrm{Li}}.$$
Because the total side-product resistance is \(R_{\mathrm{side}} = R_{\mathrm{SEI}} + R_{\mathrm{Li}}\), the terminal voltage in the final model is written as:
$$U(t) = U_p\big(y_{\mathrm{surf}}(t)\big) – U_n\big(x_{\mathrm{surf}}(t)\big) – \eta_{\mathrm{con}}(t) – \eta_{\mathrm{act}}(t) – \big(R_{\mathrm{ohm}}+R_{\mathrm{side}}(t)\big)I(t).$$
This coupled failure physical model allows the battery simulator to reproduce both the reversible voltage dynamics and the irreversible degradation trend caused by side reactions. Table below lists representative parameters for lithium plating.
| Parameter | Symbol | Value |
|---|---|---|
| Molar mass of lithium | \(M_{\mathrm{Li}}\) | \(6.94\times10^{-5}\) kg mol⁻¹ |
| Density of lithium | \(\rho_{\mathrm{Li}}\) | 534 kg m⁻³ |
| Anodic transfer coefficient | \(\alpha_a\) | 0.3 |
| Cathodic transfer coefficient | \(\alpha_c\) | 0.7 |
| Lithium ionic conductivity | \(\kappa_{\mathrm{Li}}\) | \(6\times10^{-5}\) m s⁻¹ |
2. Off-line and On-line Parameter Identification
Parameter identification is essential for making the physical model usable. Since different internal processes in an EV battery respond over different time scales, I designed a stimulus-response identification test that includes current pulses, relaxation periods, and quasi-steady charge-discharge segments. The excitation waveform was adjusted to sample every 10% state of charge across the full operating window.
2.1 Stepwise Off-line Estimation
The off-line procedure consisted of the following steps:
- Extract the end-of-relaxation voltage points to identify the open-circuit potential parameters \(y_0\), \(x_0\), \(Q_p\), and \(Q_n\).
- Measure the ohmic resistance \(R_{\mathrm{ohm}}\) by high-frequency electrical response.
- Evaluate the instantaneous voltage jump after applying current to separate the activation overpotential; then estimate the reaction polarization coefficient \(P_{\mathrm{act}}\).
- Use the end-of-charge/discharge voltage data to determine the concentration polarization coefficient \(P_{\mathrm{con}}\) and diffusion time constants \(\tau_n\), \(\tau_p\).
- After every cycle, identify the side-reaction exchange current parameters \(i_{\mathrm{SEI}}\), \(i_{\mathrm{Li}}\) using the measured \(Q_{\mathrm{Li}}\) loss.
Table represents examples of identified initial parameters for cells tested under room-temperature and low-temperature conditions. The numerical values are used as nominal starting points for the on-line estimator.
| Condition | \(y_0\) | \(x_0\) | \(Q_p\) (As) | \(Q_n\) (As) | \(R_{\mathrm{ohm}}\) (Ω) | \(P_{\mathrm{act}}\) | \(P_{\mathrm{con}}\) | \(\tau_n\) (s) |
|---|---|---|---|---|---|---|---|---|
| Room temperature | 0.102 | 0.725 | 38,066 | 42,866 | 0.0069 | 948 | 33.4 | 20.0 |
| Low temperature | 0.045 | 0.813 | 34,897 | 42,684 | 0.0069 | 16,663 | 47.4 | 291.7 |
I validated the off-line model by simulating constant-current discharge at 0.5C and 1C in three operating profiles. The mean absolute voltage error was below 15 mV, and the root-mean-square voltage error was below 26 mV over the whole discharge range. This level of accuracy indicates that the failure physical model can reproduce the voltage trajectory under different state-of-charge regions with high fidelity.
2.2 Dual Adaptive Extended Kalman Filter Framework
For real-time implementation in a EV battery management system, I developed a dual adaptive extended Kalman filter (AEKF) architecture. The original extended Kalman filter assumes fixed noise covariance matrices, which is inadequate when cell aging changes both the process dynamics and the sensor characteristics. The adaptive filter updates the process noise covariance \(Q_k\) and measurement noise covariance \(R_k\) based on the innovation sequence \(e_k\):
$$e_k = z_k – h\big(\hat{x}_{k|k-1}\big),$$
$$R_k = \frac{1}{N}\sum_{i=k-N+1}^{k} e_i e_i^T – H_k P_{k|k-1} H_k^T,$$
$$Q_k = K_k e_k e_k^T K_k^T.$$
The filter uses the standard prediction and update steps, together with the Jacobian matrices \(F_k\) and \(H_k\) of the state transition and measurement functions. In the dual architecture, the master AEKF estimates high-sensitivity parameters such as \(x_0\), \(y_0\), \(Q_p\), \(Q_n\), and \(\tau_n\). The slave AEKF then estimates low-sensitivity parameters such as \(P_{\mathrm{act}}\), \(P_{\mathrm{con}}\), and \(\tau_e\), using the outputs of the master filter as known inputs. This decomposition avoids ill-conditioning caused by simultaneous estimation of parameters with very different sensitivities.
Parameter sensitivity values are shown in the following table. The sensitivity is defined as the maximum voltage change relative to the mean voltage when the parameter varies between 50% and 150% of its nominal value.
| Parameter | Sensitivity (%) | Group |
|---|---|---|
| \(x_0\) | 12.12 | High |
| \(y_0\) | 2.47 | High |
| \(Q_p\) | 26.62 | High |
| \(Q_n\) | 12.16 | High |
| \(\tau_n\) | 1.22 | High |
| \(P_{\mathrm{act}}\) | 0.012 | Low |
| \(P_{\mathrm{con}}\) | 0.066 | Low |
| \(\tau_p\) | 0.004 | Low |
| \(\tau_e\) | 0.003 | Low |
The on-line identified parameters were compared with the off-line estimates under cycling. The trends matched well over the entire lifetime, confirming that the dual-AEKF can track the gradual drift of electrode stoichiometry and capacity. Validation through simulated voltage reproduction produced an average error below 20 mV, which is acceptable for state diagnosis in second-life EV battery management.
3. Aging Experiments and Mechanism Analysis
To accelerate degradation and mimic the transition from first-life automotive usage to second-life stationary application, I conducted cycling experiments on 8 Ah lithium iron phosphate cells. Three operating profiles were designed, covering room-temperature and low-temperature conditions with different charge-discharge rates. The schedule before and after the transition point represents the shift from normal vehicle service to milder service in second life. The experimental conditions are summarized in the table below.
| Profile | Temperature | Charge rate before transition | Discharge rate before transition | Charge rate after transition | Discharge rate after transition |
|---|---|---|---|---|---|
| 1 | 25 °C | 2C | 2C | 1C | 1C |
| 2 | 0 °C | 1C | 2C | 0.5C | 0.5C |
| 3 | 0 °C | 1C | 1C | 0.5C | 0.5C |
The characterization tests included a pulse-relaxation identification sequence, a capacity test, constant-current discharge at 0.5C, 1C, and 2C, and a dynamic stress test. The capacity test was performed every 50 cycles. All cells were cycled using a constant-current constant-voltage charge and a constant-current discharge with a resting interval in between.
3.1 Capacity Fade Characteristics
The capacity fade trajectories show distinct nonlinear features. Under room-temperature profile 1, the cell lifetime was approximately 1600 cycles. In contrast, the low-temperature cells reached the retirement threshold after about 600 cycles, indicating a roughly 2.7 times faster aging rate. The capacity trajectory under profile 1 was divided into four stages: an initial fast decline during the first 200 cycles, a temporary recovery from cycles 200–400, a slow degradation stage until cycle 1100, and a final accelerated drop near end of life. The final nonlinear drop is a signature of coupled lithium inventory depletion, polarization growth, and secondary side reactions.
Under low-temperature profiles 2 and 3, the capacity decreased in two broad phases. The initial fluctuation stage was followed by a rapid “knee” region where the capacity loss accelerates considerably. After switching to the milder second-life protocol, the capacity fading rate decreased, demonstrating that the cells still have useful life when operated under reduced current.
3.2 Evolution of Model Parameters
The identified model parameters are valuable for diagnosing failure modes. The evolution of the positive electrode capacity \(Q_p\) declined steadily under all profiles, with the low-temperature cells showing a faster decrease. Under profile 3, which uses a lower discharge rate but a longer discharge duration, \(Q_p\) decreased more significantly than under profile 2. This observation suggests that discharge duration, rather than current magnitude alone, plays a major role in low-temperature degradation of the positive electrode.
The negative electrode capacity \(Q_n\) followed a similar decreasing trend but with a smoother path. The ratio \(x_0/D_x\), which represents the fractional irreversibility of the graphite electrode, progressively increased during aging. At room temperature, the ratio remained near unity until the middle of life, then exceeded unity, indicating that the inactive lithium reservoir grew. At low temperatures, the ratio was larger from the beginning, confirming that SEI growth and lithium consumption were accelerated by the combination of low temperature and high polarization.
The reaction polarization coefficient \(P_{\mathrm{act}}\) increased by an order of magnitude at low temperature. The rise was especially strong for the cells subjected to high discharge current. Likewise, \(R_{\mathrm{ohm}}\) increased under all conditions, and the increase was more rapid at low temperature and high discharge current. These increases reflect the thickening of the SEI film, the growth of side-product layers, and the reduction of ionic conductivity within the porous electrodes.
The loss of lithium inventory \(LLI\) and loss of active material \(LAM\) were calculated from identified parameters using the following definitions:
$$LLI_N = 1 – \frac{Q_{p,N}x_{0,N} + Q_{n,N}y_{0,N}}{Q_{p,0}x_{0,0} + Q_{n,0}y_{0,0}},$$
$$LAM_N = \left(1-\frac{Q_{p,N}}{Q_{p,0}}\right) + \left(1-\frac{Q_{n,N}}{Q_{n,0}}\right).$$
Here the subscript \(N\) denotes the value after \(N\) cycles; the subscript \(0\) denotes the beginning of life. \(LLI\) represents the irreversible loss of active lithium through side reactions, while \(LAM\) combines the loss from both electrodes due to structural disintegration, particle pulverization, or isolation of active material. Table below summarizes the average growth rates of \(LLI\) and \(LAM\) before and after the transition to the second-life protocol.
| Profile | \(LLI\) rate before transition (cycle⁻¹) | \(LLI\) rate after transition (cycle⁻¹) | \(LAM\) rate before transition (cycle⁻¹) | \(LAM\) rate after transition (cycle⁻¹) |
|---|---|---|---|---|
| 1 (25 °C) | — | \(4.94 \times 10^{-5}\) average | — | \(7.72 \times 10^{-5}\) average |
| 2 (0 °C) | \(2.12 \times 10^{-4}\) | \(6.09 \times 10^{-5}\) | \(8.29 \times 10^{-4}\) | \(1.89 \times 10^{-5}\) |
| 3 (0 °C) | \(4.60 \times 10^{-4}\) | \(2.58 \times 10^{-4}\) | \(7.24 \times 10^{-4}\) | \(5.50 \times 10^{-4}\) |
After the shift to the milder second-life protocol, the rates of lithium loss and active-material loss generally decreased. Therefore, the degraded EV battery cells, although no longer suitable for high-power automotive use, can still operate safely and with a lower degradation rate in a moderate secondary application.
4. Failure Boundary Construction and Diagnosis
Different failure behaviors are controlled by different sets of physical parameters. Based on the analysis of parameter evolutions, I established the mapping between parameter groups and failure mechanisms shown below.
| Parameter set | Indicated failure behavior |
|---|---|
| \(y_0\), \(y_0/D_y\), \(LLI\) | Positive electrode lithium loss |
| \(x_0\), \(x_0/D_x\), \(LLI\) | Negative electrode lithium loss |
| \(P_{\mathrm{act}}\), \(R_{\mathrm{ohm}}\), \(LAM\) | Active material loss and polarization growth |
Because cells from the same batch exhibit a clustered evolution of parameters, I used the concepts of parameter cluster and dynamic boundary to distinguish normal aging from abnormal failure. The boundary construction pipeline consists of three steps: data generation, anomaly screening, and boundary modeling.
4.1 Data Generation through Latin Hypercube Sampling
To cover a broad but realistic parameter space, I generated synthetic parameter clusters using Latin Hypercube Sampling. For a given measured vector at cycle index \(N\), I varied the parameters \(x_0\), \(y_0\), \(Q_p\), \(Q_n\), \(P_{\mathrm{act}}\), \(P_{\mathrm{con}}\), \(R_{\mathrm{ohm}}\), and \(\tau_n\) within 80%–120% of their identified values. For each measured cycle index, I generated 2000 parameter samples. Then I simulated a constant-current discharge voltage profile and retained only those samples whose terminal voltage reached the cutoff voltage within \(\pm 50\) seconds of the actual measured discharge time. After screening, the retained parameter clusters form a realistic representation of possible cell-to-cell variations and small measurement uncertainties. Representative generation results were produced for every 100 cycles over the entire life.
4.2 Anomaly Screening with Local Outlier Factor
After constructing the synthetic clusters, I applied the Local Outlier Factor algorithm to remove points whose local density deviates from the rest of the cluster. LOF compares the local reachability density of each point with the reachability density of its \(k\) nearest neighbors. The LOF score is defined as:
$$\mathrm{LOF}(x) = \frac{\sum_{y\in N_k(x)} \mathrm{lrd}(y)}{|N_k(x)| \cdot \mathrm{lrd}(x)},$$
where \(\mathrm{lrd}(x)\) is the local reachability density:
$$\mathrm{lrd}(x) = \frac{|N_k(x)|}{\sum_{y\in N_k(x)} \mathrm{reach\_dist}_k(x,y)}.$$
A threshold is chosen such that points with an LOF score above the threshold are considered abnormal and are discarded before boundary construction. Screening with LOF improves the compactness of the boundary and avoids over-expansion caused by artificially generated extremes.
4.3 Minimum Enclosing Ball Boundary
The failure boundary is modeled as a minimum enclosing ball in the normalized multidimensional parameter space. Since different parameters have different units and orders of magnitude, all parameters were standardized and then normalized to the range \([0,1]\) using the statistics of the training cluster. The MEB formulation is
$$\min_{c,R} R^2 \quad \text{subject to} \quad \|x_i – c\|_2^2 \le R^2, \quad i=1,\ldots,N,$$
where \(x_i\) are the parameter vectors after normalization, \(c\) is the center of the ball, and \(R\) is its radius. Because the generated parameter clouds are relatively dense, I computed the center \(c\) as the arithmetic mean of all points and the radius \(R\) as the maximum distance from any point to the center:
$$c = \frac{1}{N}\sum_{i=1}^{N}x_i, \quad R = \max_{i=1,\dots,N}\|x_i-c\|.$$
This produces a compact closed boundary that encloses the normal aging paths. For each diagnosis, the observed parameter vector at every cycle is mapped into the same normalized space. If the vector lies outside the ball, the cell is diagnosed as having a failure behavior associated with the corresponding parameter set.
4.4 Results of Failure Behavior Diagnosis
I applied the failure boundary method separately to diagnose positive-electrode lithium loss, negative-electrode lithium loss, and active material loss in both room-temperature and low-temperature test groups.
For the low-temperature profile 2 group, two points in the early-mid life fell outside the boundary of positive electrode lithium loss. These periods corresponded to local capacity fluctuations and an abrupt increase in \(y_0/D_y\), indicating an abnormal lithium consumption event. Toward the end of life, several points approached the boundary even though they remained inside; the dominant cause was the accelerated rise of \(LLI\) and \(LAM\). For the low-temperature profile 3 group, one cell showed an evident transition from lithium-driven degradation to active-material-driven degradation after cycle 400. Boundary-exceeding points appeared in the active-material-loss diagnostic space at cycles 450–650 and were correlated with a rapid decrease in \(Q_p\) and \(Q_n\).
For the room-temperature profile 1 group, the positive-electrode lithium-loss boundary correctly accepted almost all points. One point outside the boundary was judged as a false positive because no physical parameter anomaly was found. For negative-electrode lithium loss, a few boundary-exceeding points appeared during the capacity recovery stage around cycle 200–300. These points can be explained by SEI reconfiguration and temporary re-activation of the graphite surface after a long rest. At the end of life, several points approached the boundary, corresponding to enhanced lithium plating and continuous SEI growth.
The overall diagnostic accuracy of the detected anomalous points was 94.1%, and the miss rate over the complete set of samples was about 8.4%. The results demonstrate that the failure boundary framework is useful for identifying abnormal evolution before a catastrophic capacity drop occurs.
5. Multi-Scale Post-Mortem Validation
To verify the failure diagnoses obtained from model parameters, I conducted post-mortem analyses on cells from different aging profiles. The validation chain includes macro-scale visual observation, meso-scale scanning electron microscopy, and micro-scale X-ray photoelectron spectroscopy.
5.1 Macro-Scale Electrode Observation
After disassembly, the positive electrodes from aged cells showed little macroscopic difference, with no obvious cracks or peeling. In contrast, the negative electrodes exhibited clear evidence of color inhomogeneity and surface wrinkles. At room temperature, the graphite electrode looked darker and had slight local folds, which is consistent with SEI film thickening and large-volume-cycled graphite particles. Under low-temperature conditions, the negative electrode displayed uneven coloring and more severe wrinkles. The middle of the electrode was darker than its edges, reflecting inhomogeneous lithium concentration and electrolyte decomposition. The low-temperature long-duration discharge profile caused the most serious morphological change, with visible gray-white deposits and curling of the electrode coating.
5.2 Scanning Electron Microscopy Analysis
SEM images of the pristine positive electrode showed well-separated particles with smooth surfaces. For the room-temperature aged cell, the positive particles were partly fractured and the binder/conductive network had degraded. In the low-temperature profile 2 cell, the cathode particles exhibited more severe cracking, and numerous deposits appeared on the particle surfaces. For the low-temperature profile 3 cell, the particle size distribution became broader; large pores and secondary cracks were observed, implying that the active material was pulverized and isolated from the conductive network.
The SEM analysis of graphite negative electrodes was even more informative. The fresh anode had well-defined flaky graphite particles with no apparent coating. In the room-temperature aged cell, the graphite particles were fractured, small fragments clogged the pores, and a dense SEI film covered the particle surfaces. The low-temperature profile 2 cell showed severe curling of graphite layers, wider interparticle gaps, and an extensive SEI layer, suggesting a strong contribution of lithium loss. The low-temperature profile 3 cell showed extensive cracking, detachment, and porous structural collapse, which is consistent with active material loss and loss of electrode integrity.
5.3 X-Ray Photoelectron Spectroscopy Analysis
XPS was used to identify the chemical constituents on the graphite surface. For the pristine anode, the deconvoluted C 1s, F 1s, Li 1s, and O 1s spectra indicate graphite carbon, a small amount of organic carbonate species, and a trace of LiF. The pristine surface chemistry is typical of a native passivation layer.
The room-temperature aged anode showed a large increase in the LiF peak and a decrease in organic C–O species. This indicates that the SEI film underwent a transformation from organic-rich to inorganic-rich composition. The increase in graphite carbon peak intensity was caused by new graphite surfaces exposed by particle fracture. No obvious lithium plating peak was detected. Thus the dominant failure behavior in the room-temperature cell was lithium loss due to SEI reconfiguration, combined with mechanical fracture of graphite particles.
The low-temperature profile 2 anode exhibited an even stronger LiF signal, while the organic carbonate peak decreased. The graphite C peak increased, indicating fresh graphite surfaces in addition to SEI growth. LiPF\(_6\) decomposition is enhanced at low temperature, and the released HF reacts to form LiF. Therefore, the dominant failure behavior is active lithium loss with significant electrolyte decomposition and graphite surface degradation.
The low-temperature profile 3 anode showed a different signature: the LiF peak increased, but the graphite C peak area shrank. A decrease in the graphite peak area suggests loss of graphitic active material. Combined with the SEM observations of severe structural collapse and detachment, this result indicates that active material loss dominates over lithium inventory loss in this cell.
5.4 Summary of Validation Results
Table below summarizes the dominant failure mechanisms identified by the multi-scale post-mortem analysis and compares them with the model-based boundary diagnosis.
| Profile | Positive electrode evidence | Negative electrode evidence | Dominant failure mode |
|---|---|---|---|
| 1 (25 °C, 2C-2C then 1C-1C) | Particle fracture, binder degradation | SEI thickening, graphite fracture, LiF increase | Lithium inventory loss |
| 2 (0 °C, 1C-2C then 0.5C-0.5C) | Surface cracks and deposits | Curling, SEI films, electrolyte decomposition | Lithium inventory loss |
| 3 (0 °C, 1C-1C then 0.5C-0.5C) | Severe cracks, large pores | Detachment, structural collapse, reduced graphite area | Active material loss |
The post-mortem observations agree with the diagnostic conclusions obtained from the failure boundary method. For the room-temperature profile, negative-electrode lithium loss was the primary threat. For low-temperature profile 2, lithium loss from the negative electrode dominated because SEI growth and electrolyte decomposition consumed active lithium. For low-temperature profile 3, active material loss became dominant because the longer discharge periods intensified structural cracking and loss of contact between particles. Thus the failure boundary method provides a credible on-line tool for failure behavior diagnosis in second-life EV battery packs.
Conclusion
I have developed a complete framework for the on-line diagnosis of failure behavior in second-life EV battery cells. First, I constructed a physics-based model that couples the main lithium intercalation/deintercalation reaction with SEI growth and lithium plating side reactions. Second, I proposed an off-line excitation-response identification procedure and a dual adaptive extended Kalman filter architecture for on-line parameter tracking. The voltage prediction errors remained below 20 mV, and the parameter trajectories were physically consistent over full-cell aging. Third, through carefully designed aging experiments under room-temperature and low-temperature profiles, I observed that the cells exhibited different capacity fade stages, and the parameters of the model evolved in a clustered, reproducible way. The evolution rates of \(LLI\) and \(LAM\) decreased after the current was reduced under the second-life protocol, confirming the practical value of battery repurposing. Fourth, I introduced a dynamic failure boundary framework based on Latin Hypercube Sampling, Local Outlier Factor screening, and a Minimum Enclosing Ball. Finally, multi-scale post-mortem analysis confirmed that the model-based diagnostics can correctly distinguish lithium inventory loss from active material loss. These findings contribute to the safe and efficient utilization of retired EV battery cells and provide a theoretical basis for next-generation battery health management systems.
