Source-linked AI summary
Infrared Hotspot-Guided Early Warning of Lithium-Ion Battery Thermal Runaway Under Mechanical Abuse
Syed Sajid Ullah, Salman Khan, Muhammad Zunair Zamir
TL;DR
Mechanical abuse can produce localized heating before conventional sensor signals become decisive, motivating earlier thermal-runaway warning. The paper uses a two-stage hotspot-guided model that distills thermal instability before multimodal prediction, achieving strong discrimination while retaining an interpretable intermediate signal.
Problem
Localized mechanical-abuse heating can precede decisive conventional sensor signals, creating a need for earlier thermal-runaway warning.
Method
A two-stage pipeline estimates thermal instability from hotspot dynamics, then fuses its score with compact multimodal features for a 20-frame warning target.
Results
Stage-I hotspot dynamics achieve ROC-AUC 0.945, while the two-stage classifier reaches Stage-II ROC-AUC 0.908.
Takeaways & Limitations
Thermal gradient rise precedes voltage-based detection by 40 frames, or 4 seconds, providing an earlier intervention window for battery management systems.
Abstract
from arXiv · showhide
Mechanical abuse can trigger thermal runaway (TR) in lithium-ion batteries through localized heat generation before sensor signals become decisive. This paper proposes a two-stage early-warning approach that estimates localized thermal instability from infrared hotspot dynamics and then fuses this instability score with mechanical, electrical, thermal, and image-intensity features for a 20-frame warning horizon. Evaluation uses repeated experiment-wise three-fold validation, with out-of-fold Stage-I scores during Stage-II training to prevent stacked-model optimism. Hotspot dynamics alone achieve Stage-I ROC-AUC 0.945, and the two-stage classifier reaches Stage-II ROC-AUC 0.908, exceeding direct multimodal fusion while preserving an interpretable intermediate instability signal. Thermal gradient rise precedes voltage-based detection by 40 frames (4 seconds) on average, enabling earlier battery management system intervention. Lead-time analysis at a fixed 0.5 threshold yields a 14.8-frame mean lead time.
I. INTRODUCTION
Mechanical abuse can create localized heating that precedes decisive conventional sensor signals. The paper addresses this gap with a leakage-controlled two-stage framework that converts hotspot dynamics into an interpretable instability score before multimodal warning prediction.
- Mechanical abuse can damage internal layers, cause short circuits, and initiate localized Joule heating that may evolve into thermal runaway.
- Conventional voltage, temperature, force, and deformation thresholds may miss spatially localized thermal responses that precede runaway.Infrared imaging captures hotspot formation, growth, gradients, entropy, and centroid motion.
- The proposed two-stage strategy estimates localized thermal instability from hotspot features, then fuses the score with mechanical, electrical, thermal, and global IR features.This separates thermal-instability detection from final warning prediction.
- The architecture uses leakage control, structured diagnostics, and an interpretable intermediate score while improving operating-threshold performance over direct multimodal fusion.
- ROC-AUC 0.945 is achieved by the compact hotspot-dynamics representation for Stage-I instability estimation.
II. RELATED WORK
Prior work provides practical sensor-based warning and spatial thermal information, but often does not separate localized instability estimation from final warning decisions. This paper combines hotspot scoring, experiment-wise leakage control, and structured reporting in a compact pipeline.
- Mechanical abuse studies examine indentation, compression, nail penetration, and impact, where localized precursors complicate early diagnosis.
- Sensor-based methods use voltage, temperature, force, strain, gas, pressure, or ultrasound, but scalar measurements can be delayed or spatially incomplete.
- Thermal imaging captures hotspot formation, spatial gradients, and propagation, while multimodal studies combine images, sensor streams, and interpretable tools.
- Prior directions do not always separate localized thermal-instability estimation from the final warning decision.The paper treats language-model-assisted safety reporting as structured classifier-output reporting rather than an independent predictor.
- The pipeline combines hotspot-derived instability scoring, experiment-wise leakage control, and structured diagnostic reporting.
- The dataset contains synchronized 10 Hz mechanical, electrical, scalar-thermal, and infrared records from cylindrical cells under indentation loading.Runaway-propagation frames are excluded from pre-runaway prediction, leaving 12,425 training rows from 194 experiments.
- Inputs are organized by mechanical, electrical, scalar-thermal, image-intensity, and hotspot-dynamics roles, with Stage II predicting TR within the next 20 frames.Event-derived variables are excluded to prevent leakage.
B. Problem Formulation
The formulation first estimates a localized thermal-instability score from hotspot dynamics alone, then augments compact multimodal features with that score to predict near-term thermal runaway.
- For each frame, the multimodal feature vector contains mechanical, electrical, scalar-thermal, image-intensity, and hotspot-dynamics features.
- Stage I estimates the localized thermal-instability state using hotspot-dynamics features only.
- The instability score ˆs_t is a probability in [0, 1] derived from the annotated thermal-instability label and observed thermal spatial patterns.
- Stage II augments compact sensor and image-intensity features with the instability score before estimating early-warning probability.
- The Stage-II target indicates whether thermal runaway occurs within the next 20 frames.
C. Baseline Classifiers
The baseline classifiers organize inputs by physical modality and compare single-stage, hotspot-focused, image-based, and multimodal configurations.
- Sensor-only uses mechanical, electrical, and scalar-thermal measurements.
- Image-all combines raw intensity statistics with hotspot-dynamics features.
- Hotspot-only uses area, centroid, gradient, entropy, growth-rate, and velocity features.
- Direct multimodal fuses all feature groups in one stage.
- The proposed two-stage configuration produces a Stage-I instability score before combining it with sensor features and four raw intensity statistics.Stage-I compares Sensor-only, Image-all, Hotspot-only, and Multimodal; Stage-II compares Sensor-only, Image-all, Direct multimodal, and the proposed design under grouped folds.
D. Two-Stage Architecture
The architecture separates localized thermal-instability estimation from final warning prediction through an interpretable bottleneck and constrained diagnostic reporting.
- Stage I receives hotspot dynamics only and outputs an instability probability.
- Stage II fuses the out-of-fold instability score with sensor variables and four global image-intensity statistics instead of directly reusing hotspot descriptors.
- The diagnostic layer reports structured evidence from the trained classifier without changing its prediction.
- LightGBM is used for both stages to model nonlinear cross-modality interactions in low-dimensional tabular data with missing hotspot coordinates.The shared hyperparameters and grouped folds support fair comparison.
E. Evaluation Metrics
Evaluation uses repeated experiment-wise three-fold cross-validation and reports discrimination, precision-recall, threshold performance, and lead time under a fixed warning criterion.
- All classifiers use experiment-wise three-fold cross-validation repeated over three random seeds.
- Mean ROC-AUC with min-max ranges across folds is reported for both stages.
- Average precision, F1, and balanced accuracy are reported at the default threshold of 0.5.
- Lead time uses the same 0.5 threshold, requiring a warning to cross the threshold before thermal runaway and within the 20-frame horizon.
IV. RESULTS AND DISCUSSION
Results examine modality contributions, thermal-signal timing, and threshold-based lead time across the two-stage early-warning framework.
- Thermal evidence: In Experiment 84, surface temperature rises from 31°C to 230°C while voltage remains at 4.0 V and hotspot area shows no appreciable growth.
- Thermal evidence: In Experiment 22, voltage declines from 4.0 V to 1.8 V while hotspot area grows consistently.
- Temporal warning: Thermal gradient rise provides a mean 40.3-frame lead before runaway, compared with 21.0 frames for voltage below 3.5 V.
- Temporal warning: Thermal gradient detects pre-runaway anomalies in 143 of 199 experiments (71.9%), versus 74 for the voltage threshold.Coverage is broader but not universal.
- Stage-I contributions: Hotspot-only reaches Stage-I ROC-AUC 0.945, while area-based cues are strongest at 0.942 and centroid features follow at 0.923.
B. Stage II: Early-Warning Prediction
Stage II combines the learned instability score with multimodal measurements for early-warning prediction. The two-stage classifier outperforms direct multimodal fusion on ROC-AUC and threshold-level metrics.
- 0.908 ROC-AUC is achieved by the full two-stage classifier, versus 0.903 for direct multimodal fusion.The two-stage model also reaches F1 0.752 versus 0.738 and BAcc 0.836 versus 0.824.
- Mechanical variables are the strongest single physical modality at ROC-AUC 0.871, ahead of electrical, scalar-thermal, and thermal-image pathways.The corresponding single-modality ROC-AUC values are 0.706, 0.657, and 0.655, respectively.
- Thermal signals become abnormal before voltage reaches a critical reading in the multimodal trajectories.The figure marks stable, precursor, localized, and runaway stages, with a dashed line indicating TR onset.
- Stage-II ROC and precision-recall curves compare the proposed classifier with alternative classifier configurations.AUC and AP values are displayed in the figure legend.
C. Ablation Study and Diagnostic Reporting
Ablations show that the explicit instability score and compact global image context jointly support Stage-II performance. Diagnostic analyses identify hotspot-driven contributions, while threshold 0.5 provides near-optimal operating behavior within the evaluated dataset.
- Ablation Study: 0.896 ROC-AUC results when raw intensity statistics are removed, compared with 0.905 without the Stage-I score and 0.908 for the complete model.The ablation identifies the combination of explicit instability scoring and global image context as strongest.
- Diagnostic Reporting: Growth rate, thermal gradient, and area have positive SHAP contributions toward Stage-I instability detection.In Stage II, the injected instability score is dominant, followed by compact image-intensity and voltage-related cues.
- Threshold Analysis: 14.8 frames is the mean lead time for 183 valid in-horizon warnings at threshold 0.5.Other detections occur beyond the 20-frame horizon; the best mean F1 at threshold 0.55 improves over 0.5 by less than 0.001.
- Scope: The evaluation remains dataset-specific, using 194 experiments with pre-runaway frames, while the earliest handcrafted cue appears in 143 of 199 experiments.
- Diagnostic Reporting: The diagnostic reporting module verbalizes structured evidence from the trained warning model without altering the classifier decision.
V. CONCLUSION
The paper presents a two-stage hotspot-aware framework for warning of mechanically induced lithium-ion battery thermal runaway. It distills hotspot dynamics into an interpretable instability score before multimodal fusion, with cross-setting validation identified as future work.
- 0.945 Stage-I ROC-AUC and 0.908 Stage-II ROC-AUC support hotspot-aware instability estimation followed by compact multimodal fusion.Thermal signatures precede voltage collapse by up to 40 frames on average.
- The instability score remains the dominant Stage-II driver according to SHAP analysis.
- Future work should validate the model across additional abuse modes, cell chemistries, and pack-level propagation settings.Deployment-specific threshold calibration and end-to-end thermal-video alternatives are also identified for study.