Deployable Self-Referenced Features for Cross-Cell Battery State-of-Health Estimation: Randomized Pairing Controls and Benchmarking

Abstract

Battery health estimation depends on feature references as well as learners. Changing one column’s normalization moves root-mean-square error (RMSE) by up to 108%. We construct a self-referenced lift (SRL) from each cell’s earliest retained record. Under the offline author protocol, learners with the full history stack improve on 9 of 11 units after Holm correction, with none significantly worse and two units per backbone not significantly different. Relative-change medians over all 11 units are −16.8% for the frozen foundation model and −19.1% for gradient boosting, including learner effects. Under a strict online reader excluding target-future and label information, compact SRL instead yields +4.3% for gradient boosting (5 of 11 improve) and +10.4% for the frozen foundation model (3 of 11 improve): no overall online accuracy gain. Legacy offline pairing controls favour self-pairing (p=0.0099, 100 derangements). A companion benchmark with implicit self-comparison improves in 6 of 6 batches with per-batch selection (5 of 6 with a fixed configuration); an unseen-library legacy diagnostic improves by 22–33%. Deployability concerns feature computability, not a strict-online accuracy advantage.