Abstract
The identification of the "knee" in lithium-ion battery capacity fade—the point where degradation accelerates—is critical for optimizing battery life and safety. This paper presents an unsupervised framework for identifying "knee-onset," a precursor to the knee, using time-series segmentation. By applying the Scalable Time series Anytime Matrix Profile (STAMP) algorithm to degradation curvature, we identify a previously uncharacterized oscillatory transition phase. Our method is validated across LFP, NMC, and synthetic datasets, demonstrating that knee-onset can be identified hundreds of cycles before the knee occurs, enabling early end-of-life prediction with high accuracy.
Introduction
Lithium-ion batteries often exhibit a non-linear degradation profile characterized by a "knee," where the capacity fade rate increases abruptly. For electric vehicle (EV) manufacturers and second-life battery recyclers, identifying this point is essential for effective battery management. However, existing methods often rely on rigid parametric models or data-intensive supervised learning, which may fail to generalize across different chemistries or require excessive labeled data. This study introduces an unsupervised approach that treats battery aging as a time-series segmentation problem, focusing on the "knee-onset" as a more proactive health indicator.
Methodology
The proposed unsupervised pipeline consists of four primary stages:
- Denoising: Raw capacity data is smoothed using a Savitzky-Golay filter to preserve the underlying degradation signal while removing noise.
- Curvature Approximation: Instead of direct capacity analysis, the "approximated curvature" is derived from successive differences, highlighting the rate of change in the degradation rate.
- Matrix Profile & Arc Curves: The STAMP algorithm is employed to find self-similarity in the curvature data. This leads to the generation of a Corrected Arc Curve (CAC), which serves as a robust regime change detector.
- Regime Extraction: The algorithm segments the battery's life into three distinct states: stable degradation (State 1), an oscillatory transition phase (State 2, or knee-onset), and accelerated fade (State 3, or the knee).
Datasets and Validation
The framework's generalizability was tested across three diverse data sources:
- Toyota Research Institute (TRI): 169 LFP/graphite cells, where knees are primarily driven by lithium plating.
- Sandia National Lab (SNL): 22 NMC/graphite cells, where knees result from resistance growth (SEI layer thickening).
- Synthetic Dataset: Generated using the PyBaMM library (Doyle-Fuller-Newman model) to simulate knees caused by particle cracking.
To validate the segmentation algorithm's ability to identify transition points, synthetic degradation curves were modeled using specific parameters to define the intercept, slopes, and the ground-truth knee-onset point.
Results and Discussion
A significant finding of this research is the "oscillatory phenomenon"—a period of high curvature fluctuation that consistently precedes the knee. This transition state appears to be a universal precursor to failure across different chemical mechanisms. Identifying the knee-onset provides a warning approximately 280 to 323 cycles earlier than traditional knee detection methods.
The identified knee-onset shows a near-perfect linear correlation with the battery's actual End of Life (EoL), with Pearson’s correlation coefficients ($r$) ranging from 0.992 to 1.0 for LFP cells. Furthermore, by training a Gradient Boosting Regression Tree (GBRT) on the first 30 cycles of data, the model predicted knee-onset with an RMSE of 59.2 cycles. This early prediction capability translates to a reduction of roughly 269 hours of experimental testing time per cell in the TRI dataset.
Conclusion
This study demonstrates that unsupervised time-series segmentation of degradation curvature can effectively identify critical transition points in battery life. The discovery of the oscillatory knee-onset phase provides a significant lead time for battery health management. Future research will focus on validating this method against dynamic driving profiles and adapting it for real-time estimation using field data from electric vehicles.
Summary updated