Cross-Family Transferability of Multimodal Sensor Information for Short-Horizon Forecasting of Mechanically Induced Lithium-Ion Battery Thermal Runaway
- 1School of Energy and Mechanical Engineering, Dezhou University, Dezhou 253023, China
- 2School of Mechanical Engineering, Dalian Jiaotong University, Dalian 116028, China
- 3School of Materials Science and Engineering, Dalian University of Technology, Dalian 116024, China
Abstract
Additional sensor modalities may improve short-horizon lithium-ion battery thermal-runaway (TR) forecasting within familiar experimental domains, but whether that benefit persists in unseen cell/source families is unclear. This study evaluated cross-family transferability using 67 mechanically induced TR experiments from six cell/source families. Past 3-s histories of temperature (T), terminal voltage (V), and penetrator load/force (F) were used to forecast 1, 2, and 5 s temperature increments. At the focal 2 s pre-onset horizon, Extra Trees utilities for T + V, T + F, and T + V + F were 11.2%, 17.5%, and 12.4% under experiment-grouped random validation but −3.5%, −6.3%, and −14.3% under leave-one-family-out validation. Removing absolute V(t) reduced the voltage transfer penalty point estimate from 14.7 to 1.8 percentage points; further relative scaling produced −0.8 percentage points but no consistent incremental benefit. Across six families, joint voltage–force feature discrepancy showed a conditional, small-sample association with transfer penalty under the dynamic maximum target, (Spearman ρ = 1.00, exact p = 0.0028); this association attenuated when the median thermocouple (TC) temperature or first thermocouple channel (TC1) was used. Mechanical sensor transfer estimates likewise changed materially with target construction. Within this mechanically induced TR dataset, these results show that apparent multimodal gains depend on validation design, sensor representation, and target temperature definition. Cross-family claims should therefore be supported by family domain holdout validation and harmonized input and target representations. The focal pre-onset analysis is short-horizon regression conditional on a retrospectively defined window (−10 ≤ τ < −2 s relative to operational TR onset), not an autonomous prospective warning trigger.
Related
- Cross-exposure transferability and failure boundaries of self-supervised DNN denoising for AR-HAXPES depth profiling
- A predictive transferability index for crystal plasticity framework selection: Variational derivation, analytical bounds, and FFT-based plausibility assessment
- Upcycling Waste Bamboo Residues into Mechanically Enhanced TOCNF/Quaternized Cellulose Composite Films
- Laminated structure design accommodates thermal expansion mismatch for mechanically and thermally dual-stable thermoelectric contacts
- Rapid Aryne-Enabled Aryl Functionalization via Mechanically Assisted Microscale Mixing
BibTeX
@article{li2026crossfamily,
title = {{Cross-Family Transferability of Multimodal Sensor Information for Short-Horizon Forecasting of Mechanically Induced Lithium-Ion Battery Thermal Runaway}},
author = {Yuanzhe Li and Ningning Wei and Qiang Li and Binguang Jia and Minghai Li and Lei Huo},
journal = {Batteries},
year = {2026},
month = oct,
doi = {10.3390/batteries12110445},
url = {https://www.mdpi.com/2313-0105/12/11/445},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Cross-Family Transferability of Multimodal Sensor Information for Short-Horizon Forecasting of Mechanically Induced Lithium-Ion Battery Thermal Runaway
AU - Yuanzhe Li
AU - Ningning Wei
AU - Qiang Li
AU - Binguang Jia
AU - Minghai Li
AU - Lei Huo
PY - 2026
DA - 2026/10/30
JO - Batteries
DO - 10.3390/batteries12110445
UR - https://www.mdpi.com/2313-0105/12/11/445
AB - Additional sensor modalities may improve short-horizon lithium-ion battery thermal-runaway (TR) forecasting within familiar experimental domains, but whether that benefit persists in unseen cell/source families is unclear. This study evaluated cross-family transferability using 67 mechanically induced TR experiments from six cell/source families. Past 3-s histories of temperature (T), terminal voltage (V), and penetrator load/force (F) were used to forecast 1, 2, and 5 s temperature increments. At the focal 2 s pre-onset horizon, Extra Trees utilities for T + V, T + F, and T + V + F were 11.2%, 17.5%, and 12.4% under experiment-grouped random validation but −3.5%, −6.3%, and −14.3% under leave-one-family-out validation. Removing absolute V(t) reduced the voltage transfer penalty point estimate from 14.7 to 1.8 percentage points; further relative scaling produced −0.8 percentage points but no consistent incremental benefit. Across six families, joint voltage–force feature discrepancy showed a conditional, small-sample association with transfer penalty under the dynamic maximum target, (Spearman ρ = 1.00, exact p = 0.0028); this association attenuated when the median thermocouple (TC) temperature or first thermocouple channel (TC1) was used. Mechanical sensor transfer estimates likewise changed materially with target construction. Within this mechanically induced TR dataset, these results show that apparent multimodal gains depend on validation design, sensor representation, and target temperature definition. Cross-family claims should therefore be supported by family domain holdout validation and harmonized input and target representations. The focal pre-onset analysis is short-horizon regression conditional on a retrospectively defined window (−10 ≤ τ < −2 s relative to operational TR onset), not an autonomous prospective warning trigger.
SN - 2313-0105
ER -