Technical Summary: Health Feature Extraction from Battery Energy Storage System Field Fault Data

Problem Statement

Monitoring the health of lithium-ion battery modules in grid-connected Battery Energy Storage Systems (BESS) is critical for preventing catastrophic failures. However, assessing cell health in the field presents significant challenges distinct from controlled laboratory environments:

1. Variable Operating Conditions: Real-world systems experience fluctuating C-rates, voltage windows, and temperatures. These variations directly confound health features, making it difficult to distinguish between changes caused by degradation and those caused by operational dynamics.
2. Sparse Sensing: Commercial modules typically lack sensors for every individual cell. Voltage and current are measured at the parallel-group level, and temperature sensors are sparse. This prevents direct observation of individual cell states, obscuring subtle fault signatures.
3. Unknown Fault Mechanisms: Unlike laboratory studies where abuse conditions are induced, field faults have unknown root causes, complicating the identification of specific fault indicators.
4. Data Scarcity: Fault data is rare, and existing public datasets often lack parallel-connected cell configurations or rely exclusively on internal resistance as a health indicator.

Methodology

The authors present a framework to extract and calibrate health features from operational data to identify discriminative features for faulty parallel-connected cell groups. The methodology consists of three main stages:

1. Data Processing and Feature Extraction

Using 1 Hz operational data from 25 commercial BESS modules (each containing 14 series-connected parallel groups), the authors extracted three categories of health features for each parallel group:

• Capacity ($\hat{C}$): Estimated during constant-power discharge events using Coulomb counting. State of Charge (SOC) boundaries were determined by mapping relaxed terminal voltages to a fresh-cell Open-Circuit Voltage (OCV) curve.
• Internal Resistance ($\hat{R}$): Estimated from voltage relaxation behavior following charging events. Resistance was calculated at various time steps ($\tau \in \{1s, 5, 10, 15, 20, 25 \text{ min}\}$) to capture ohmic, charge-transfer, and diffusion resistances.
• Differential Voltage Analysis (DVA) Features: Extracted from $dV/dQ$ curves during discharge. Features included the height and voltage location of the High Voltage (3.7–3.9 V) and Low Voltage (3.45–3.65 V) peaks. Specific algorithms (OCP-informed identification and Savitzky-Golay filtering) were used to mitigate sensor noise.

2. Gaussian Process (GP) Regression for Calibration

To decouple the effects of variable operating conditions from the underlying degradation trend, the authors employed a GP regression framework.

• Model Structure: Each health feature $y$ was modeled as the sum of two independent GPs: $f_{op}(x)$, capturing sensitivity to operating conditions (e.g., C-rate, voltage, current), and $f_{deg}(t)$, capturing the time-dependent degradation trajectory.
• Calibration: By isolating the $f_{deg}(t)$ component, the method produces "calibrated" features that represent the true aging trend regardless of the specific operating conditions at the time of measurement.
• Normalization: To enable cross-module comparison, calibrated features were normalized relative to the mean and standard deviation of the other 13 groups within the same module.

3. Statistical Evaluation

The authors evaluated the ability of these calibrated features to distinguish the single confirmed faulty group from the 13 non-faulty groups in each of the 25 modules.

• Statistical Tests: Mann–Whitney U tests were used to determine if the distributions of faulty vs. non-faulty groups differed significantly ($p < 0.05$).
• Discriminative Utility: The Area Under the Receiver Operating Characteristic Curve (AUC) was calculated to quantify the practical detection capability of each feature.

Key Results

The study utilized a dataset comprising 25 faulty and 325 non-faulty cell groups, with over 183 million data rows collected over 1–6 months prior to fault confirmation.

• Significant Separation:

  • Capacity: Both the median capacity values and the capacity degradation rate (slope) showed statistically significant differences between faulty and non-faulty groups ($p < 0.05$). Faulty groups exhibited lower capacity and accelerated degradation as early as 15–30 days before the fault.
  • DVA Peak Heights: The heights of both High V and Low V $dV/dQ$ peaks were significantly lower for faulty groups ($p < 0.05$).
  • Note: DVA peak locations and capacity slopes at certain time windows did not always show significance, but the primary metrics (capacity and peak heights) did.

• Lack of Separation:

  • Internal Resistance: Neither the ohmic resistance ($\tau=1s$) nor the charge-transfer/diffusion resistances ($\tau > 1s$) showed statistically significant differences between faulty and non-faulty groups ($p > 0.05$).

• Detection Performance:

  • Despite statistical significance, the discriminative power (AUC) of all features was modest. The best-performing feature, capacity degradation rate, achieved an AUC of 0.671. DVA peak heights and median capacity values ranged from 0.614 to 0.64.
  • Resistance features performed near random guessing (AUC < 0.60).
  • The authors attribute this to substantial distributional overlap between faulty and non-faulty populations, likely due to heterogeneous fault mechanisms and the masking effect of parallel connections.

Hypothesized Failure Mechanisms

Based on the observed feature signatures, the authors hypothesize that the faults in this dataset are consistent with:

• Cell-to-cell imbalances and inhomogeneities: Indicated by reduced $dV/dQ$ peak heights, which suggest broader phase transitions and uneven current distribution.
• Local Lithium Plating: Indicated by lower capacity and accelerated degradation rates, consistent with irreversible loss of lithium inventory (LLI).

The lack of resistance deviation suggests that the faults were not primarily characterized by significant resistance increases (e.g., connector corrosion or internal shorts), or that the parallel architecture masked the resistance increase of a single failing cell.

Significance and Claims

The paper claims the following contributions and significance:

1. Framework for Field Data: It demonstrates a robust framework for extracting and calibrating health features from real-world BESS operational data, successfully decoupling degradation trends from variable operating conditions using Gaussian Processes.
2. Challenge to Resistance-Centric Detection: The findings challenge the exclusive reliance on internal resistance for fault detection in parallel-connected modules. The study shows that in this specific dataset, resistance was not a primary indicator of fault, whereas capacity and DVA features were.
3. Understanding Parallel Fault Dynamics: By analyzing parallel-connected groups, the work addresses a gap in existing literature (which often focuses on series-only cells) and highlights the "masking effect" where parallel connections can hide individual cell degradation in group-level resistance measurements.
4. Realistic Expectations: The authors modestly conclude that while statistically significant differences exist, the substantial overlap in feature distributions highlights the inherent difficulty of using individual electrochemical features for robust fault detection in the field without incurring high false-positive rates. The work advances the understanding of fault behavior but does not claim to have solved the detection problem entirely.