Published January 21, 2026 | Version v1

CRYPTOGRAPHICALLY-ATTESTED BATTERY STATE-OF-HEALTH ESTIMATION ON STOCK CONSUMER ANDROID VIA CHARGING-CURVE REPRESENTATION LEARNING

Authors/Creators

Description

The secondary smartphone market is undermined by “lemon” devices with degraded batteries masquerading as
healthy. We present TVO (Temporal Voltage Oracle), a system for State-of-Health (SoH) estimation that
encodes charging-curve telemetry as Gramian Angular Field (GAF) images on stock, unrooted consumer
Android and estimates SoH via a deployable constrained-binary estimator pipeline. The currently deployed v1 is
a closed-form analytic model; a gradient-boosted regressor exported to ONNX is evaluated on-device as the
next production candidate; a full CNN-backbone path is validated offline pending ONNX Runtime vendoring.
Our primary contribution is PIACS-RDA (Physics-Informed Aging-Curve Synthesis with Real-Device
Anchoring), a training-data methodology that augments healthy charging captures from 347 physical devices
into a robust, labeled aging corpus by applying electrochemical degradation transforms grounded in real
measured aging patterns. All devices were tested at the Retronics Lab facility; ground-truth SoH was established
via calibrated benchtop cyclers for every device using a standardised cycler protocol (0.5C CC-CV, IEC 61960
capacity definition).
We further contribute a degradation-knob prediction layer that estimates each of four synthetic aging parameters
separately and composes them, supporting interpretable per-factor attribution rather than an opaque scalar
verdict. The four predicted knobs are the same parameters used during PIACS-RDA data augmentation; they are
validated against their synthesis targets, not against independently measured electrochemical mechanisms.
On group-aware five-fold cross-validation with device-level grouping, preventing data leakage from augmented
variants, the gradient-boosted candidate achieves MAPE 4.21 ± 0.09%; the deployed closed-form analytic
estimator achieves MAPE 8.24 ± 0.70%; an offline-evaluated CNN backbone achieves MAPE 3.85 ± 0.07%.
All MAPE values are computed on original real device traces validated against independent cycler ground truth;
synthetic augmented samples are used for training only and never appear in test folds. Finally, we provide a
TEE-rooted cryptographic envelope that signs the SoH verdict, ensuring the integrity of health disclosures in
refurbishment workflows.

Files

CRYPTOGRAPHICALLY-JAN2026-51.pdf

Files (768.9 kB)

Name Size
md5:25ca03f02f6a3ebf1137ab997d03fc5e
768.9 kB Preview Download

Additional details