New Algorithm Cuts Battery Inconsistency Analysis Time by 96.7%
In the race to improve electric vehicle (EV) reliability and longevity, a critical yet often overlooked challenge lies beneath the hood: battery pack inconsistency. As EVs proliferate globally, ensuring the health and uniformity of hundreds of lithium-ion cells within each pack has become a linchpin for safety, performance, and resale value. Now, a team of researchers from Yanshan University and Geely Holding Group has unveiled a breakthrough method that slashes computational time for assessing battery inconsistency by over 96%—without sacrificing accuracy.
The innovation, detailed in a recent study published in Acta Metrologica Sinica, combines an adaptive downsampling technique with an advanced time-series clustering algorithm to evaluate voltage data from real-world EV operations. Unlike traditional approaches that rely on internal battery parameters—such as state of charge (SOC) or internal resistance—which are difficult to measure in real time, this new method leverages only the voltage signals already captured by standard battery management systems (BMS). This makes it immediately deployable across existing EV fleets.
At the heart of the solution is a refined version of the Largest-Triangle-Three-Buckets (LTTB) algorithm, dubbed “adaptive LTTB.” While conventional LTTB reduces data volume by preserving the visual shape of time-series curves, it applies a fixed compression ratio across all segments. This can distort critical features—especially during charging phases, where voltage plateaus and rapid rises carry vital diagnostic information. The adaptive LTTB, by contrast, dynamically allocates sampling points based on the local complexity of the voltage curve. It preserves high resolution during volatile transitions and reduces it during stable periods, all while automatically determining the optimal compression ratio for each charge-discharge cycle.
“We observed that existing downsampling methods either oversimplify the data or fail to adjust to the unique characteristics of different driving and charging patterns,” said Wu Fenghe, lead author and professor at Yanshan University’s School of Mechanical Engineering. “Our adaptive approach ensures that the essential shape features of voltage sequences—particularly during the two distinct plateau phases of lithium-ion charging—are retained, even after aggressive data reduction.”
This refined data is then fed into a time-series clustering pipeline based on DTW-DBA-Means, a sophisticated algorithm that outperforms mainstream alternatives like k-Shape in capturing the nuanced dynamics of battery behavior. DTW (Dynamic Time Warping) measures similarity between voltage sequences by aligning peaks and troughs—even if they occur at slightly different times—while DBA (DTW Barycenter Averaging) computes a meaningful cluster center that reflects the shared temporal structure of similar cells.
The result is a robust inconsistency metric: the silhouette coefficient. Ranging from -1 to 1, this statistical measure quantifies how well individual battery cells group together based on their voltage trajectories. A higher silhouette coefficient indicates greater divergence among cell behaviors—i.e., worse inconsistency. Crucially, the researchers validated this metric against a ground-truth proxy: the dispersion of actual discharge capacities across cells, calculated from real-world current and voltage data during deep-discharge events (70% depth of discharge). The correlation was strong, confirming that voltage-based clustering reliably reflects underlying electrochemical divergence.
The practical implications are substantial. In tests using nine months of operational data from a real EV—comprising 48 full charge-discharge cycles—the new method achieved a silhouette coefficient within the normal range of 0.111 to 0.292, consistent with a vehicle showing no BMS warnings or performance issues. More impressively, the entire analysis, which would have taken over 17 hours (62,850 seconds) using raw, unprocessed data with DTW-DBA-Means alone, was completed in just under 35 minutes (2,079 seconds) when paired with adaptive LTTB—a 96.7% reduction in runtime.
This efficiency gain is not merely a technical footnote. For automakers and fleet operators managing thousands of vehicles, rapid and scalable battery health assessment is essential for predictive maintenance, warranty management, and second-life battery repurposing. Current BMS software often lacks the computational bandwidth to run complex clustering algorithms in real time or even daily. By compressing data intelligently before analysis, the adaptive LTTB method bridges this gap, enabling high-fidelity diagnostics on standard onboard hardware or in cloud-based analytics platforms without costly upgrades.
Geely Holding Group, a co-affiliate of the research team, is already evaluating the integration of this technique into its next-generation BMS architecture. “Battery inconsistency is a silent killer,” said Zhang Zhengzhu, a senior engineer at Geely and co-author of the study. “It doesn’t trigger immediate faults, but it accelerates degradation, reduces usable range, and increases the risk of thermal runaway over time. Early detection through efficient, data-driven methods like this one allows us to intervene before problems escalate—whether through cell balancing, module replacement, or usage recommendations to the driver.”
The method also aligns with global trends toward data-centric EV diagnostics. As vehicles become increasingly connected, terabytes of telemetry data stream from BMS units every day. However, much of this data is discarded or heavily averaged due to storage and processing constraints. Techniques that extract maximum insight from minimal data—without losing diagnostic fidelity—are therefore in high demand. The adaptive LTTB approach exemplifies this philosophy: it doesn’t just reduce data; it reduces it intelligently, preserving the signals that matter most for health assessment.
Moreover, the framework is inherently scalable. While the current study focused on voltage data from a single vehicle model, the algorithm is agnostic to battery chemistry, pack architecture, or driving conditions. With minor calibration, it could be applied to commercial EVs, e-buses, or even grid-scale energy storage systems—all of which suffer from cell-to-cell variability over time.
Critically, the research adheres to rigorous scientific standards. The team used real-world operational data compliant with China’s GB/T 32960-2016 standard for EV telematics, ensuring relevance to actual driving scenarios rather than idealized lab conditions. They also conducted ablation studies, comparing their full pipeline against variants using standard LTTB or dynamic LTTB, and benchmarked DTW-DBA-Means against k-Shape. In every case, the proposed method demonstrated superior accuracy and efficiency.
For investors and industry analysts, the development underscores a broader shift in EV technology: the move from hardware-centric innovation to software-defined intelligence. While cell chemistry and thermal management remain vital, the ability to understand and act on battery data in real time is becoming a key differentiator. Companies that master this domain will not only improve vehicle safety and longevity but also unlock new revenue streams through battery-as-a-service models, predictive maintenance contracts, and certified pre-owned EV programs with verified battery health reports.
Looking ahead, the researchers plan to extend the method to incorporate temperature and current data, which could further refine inconsistency detection—especially in extreme climates or high-performance driving scenarios. They are also exploring integration with machine learning models that predict remaining useful life based on evolving silhouette coefficients over time.
In an industry where milliseconds can mean the difference between safe operation and catastrophic failure, a 96.7% reduction in diagnostic latency is more than an engineering triumph—it’s a potential game-changer for EV reliability worldwide.
Author: Wu Fenghe¹, Chai Haining¹, Zhang Zhengzhu¹,², Zhang Ning¹, Wang Zhengming², Jiang Zhanpeng¹, Guo Baosu¹
Affiliations:
¹ School of Mechanical Engineering, Yanshan University, Qinhuangdao, Hebei 066004, China
² Geely Holding Group, Hangzhou, Zhejiang 310051, China
Journal: Acta Metrologica Sinica
DOI: 10.3969/j.issn.1000-1158.2024.06.0890