TL;DR
This paper introduces a novel data-driven framework to predict the risk of thermal runaway (TR) in lithium-ion batteries as they age. By fusing multiple degradation indicators using Physics-Informed Neural Networks (PINNs) and employing a stochastic Wiener process model, the authors provide a dynamic, probabilistic assessment of TR risk throughout a battery's lifecycle, crucial for enhancing safety in electric vehicles and other applications.
What Problem Does This Paper Address?
Lithium-ion batteries are essential for modern electric vehicles and other technologies, but their safety, particularly concerning thermal runaway (TR), remains a significant hurdle to widespread adoption. While much research has focused on the physical mechanisms of TR and developing safer materials, there's a critical gap in understanding how the gradual degradation of batteries over their operational life influences their susceptibility to TR. Existing methods often struggle to dynamically predict this evolving risk, especially under real-world, variable operating conditions. This paper aims to bridge this gap by developing a prognostic model that quantifies TR risk based on observable degradation patterns.
Methods
The proposed framework integrates data-driven and physics-informed approaches.
- Multivariate Degradation Information Fusion: The authors first identify eight degradation-sensitive parameters from charging curve segments, including mean current, current standard deviation, and charging time. These parameters are then fused using Physics-Informed Neural Networks (PINNs). The PINNs embed electrochemical-thermal coupling mechanisms as constraints, allowing for the inference of unobservable internal battery states (like SEI film growth) from readily available external data. This fusion process yields a composite degradation index.
- Stochastic Process Modeling: Recognizing that both battery degradation and external TR-triggering stresses are inherently stochastic, the paper employs a Wiener process-based random failure threshold model. The degradation trajectory of the composite index (related to the self-heating onset temperature, T1) is linearized and modeled as a Wiener process. The model then accounts for the random nature of the failure threshold (L) by assuming it follows a normal distribution. This allows for the probabilistic prediction of the time to thermal runaway.
Datasets
The study utilizes the Open Access XJTU Battery Dataset, specifically focusing on 18,650 nickel-cobalt-manganese (NCM) lithium-ion batteries. The dataset includes charging and discharging data under specific test conditions (2C charging, 1.0C discharging). The authors also reference the work of Ren et al. [40] for validating the T1-SOH relationship, which is crucial for their model.
Key Insights and Ideas
- Degradation as a TR Precursor: The paper strongly emphasizes that battery aging isn't just about capacity loss; it directly impacts thermal stability and increases TR risk. This shifts the focus from reactive safety measures to proactive, degradation-aware safety management.
- PINNs for Unobservable States: Using PINNs to fuse multiple degradation features allows for the inference of critical internal states (like SEI growth) that are not directly measurable but significantly influence thermal stability. This is a powerful way to leverage available data for deeper insights.
- Stochasticity is Key: Acknowledging and modeling the inherent randomness in both battery degradation and external trigger events is crucial for realistic TR risk prediction. A deterministic approach would likely underestimate or overestimate risk in real-world scenarios.
- T1 as a Central Indicator: The self-heating onset temperature (T1) is identified as a key indicator of thermal stability, and its degradation trajectory is central to the risk prediction model.
Key Results
- The proposed framework successfully fuses eight current-related degradation features using PINNs to estimate the State of Health (SOH).
- An approximate relationship between SOH and T1 (thermal stability indicator) was established and used to derive a T1 degradation curve.
- A time-scale transformation (t = Cycle²) was applied to linearize the T1 degradation, achieving a high linear relationship (R² = 0.977).
- The Wiener process model, with estimated drift (μ = -6.177 × 10⁻⁵) and diffusion (σ = 0.03524) coefficients, was used to model this linearized degradation.
- The thermal runaway risk probability curve, generated through Monte Carlo simulation, shows an S-shaped growth. The risk remains low for the first ~500 cycles, then increases non-linearly, exceeding 10% around 1000 cycles. This highlights a critical window for intervention.
Limitations and Future Work
- T1 Measurement Dependency: The study acknowledges a limitation: the XJTU dataset did not include direct measurements of the thermal runaway characteristic parameter (T1). The risk modeling relied on an offline-calibrated T1-SOH relationship from other research. The model's accuracy is thus dependent on the precision of this relationship.
- Future Work Suggestions:
- Conduct direct T1 data measurement experiments for various battery chemistries and operating conditions.
- Extend the methodology to different battery chemistries.
- Develop online updating mechanisms for real-time parameter identification and dynamic risk prediction.
Summary updated