A Noise-Aware Calibrated Multi-Scale Voltage-Residual Method for Internal Short-Circuit Detection in Lithium-Ion Batteries

Detecting internal short-circuit (ISC) conditions in lithium-ion batteries is difficult because fault-induced voltage deviations are often masked by load transients, state-of-charge variation, model mismatch, and measurement noise. This study proposes a noise-aware calibrated multi-scale voltage-residual method for ISC detection. A healthy-voltage predictor is first trained using normal operating data to construct a reference trajectory, and measured-minus-predicted voltage residuals are then calibrated with train-normal statistics. Multi-scale and current-coupled residual descriptors are extracted to characterize both local deviations and persistent voltage mismatch. The experimental protocol uses experiment-record/source-file-grouped training, validation, and locked test partitions; all threshold selection and decision refinement are performed using validation data only. On the clean locked test set under this protocol, the final detector achieved F1 = 0.8100, ROC-AUC = 0.8902, PR-AUC = 0.9346, and balanced accuracy = 0.8026. Its recall was 0.7393, corresponding to 3005 missed ISC windows among 11,525 ISC-labeled test windows, i.e., a false-negative (miss) rate of 0.2607 (26.07%). Relative to the calibrated multi-scale baseline (F1 = 0.8291, specificity = 0.5504, FAR = 0.4496), the final robust rule trades some F1 for substantially higher specificity (0.8659) and a lower FAR (0.1341); its principal incremental benefit is therefore false-alarm control rather than uniform superiority across metrics. Under the fixed-rule artificial Gaussian voltage-noise stress test, ROC-AUC is 0.8432 with FAR = 0.1149 at the 0.10% nominal perturbation level, but ROC-AUC decreases to 0.7021 at 0.50% and to 0.6249 at 2.00%, with FAR increasing to 0.5235 and 0.7469, respectively. The noise evidence therefore supports low-noise false-alarm mitigation under the specified Gaussian protocol, not broad robustness to substantial or realistic sensor noise. These results support experiment-record/source-file-grouped offline evaluation of the proposed residual formulation. The supervised endpoint is binary normal-versus-resistor-emulated ISC classification; short-circuit resistance Rsc is not a prediction target. The results do not by themselves establish cell-level transferability, realistic sensor robustness, quantitative Rsc estimation, or online deployment. Because the public release does not provide a verified pointwise ISC-onset annotation, detection delay, pre-onset false-alarm rate, warning lead time, and horizon-dependent performance cannot be computed; accordingly, this study does not claim early detection.