Noise-robust differential thermal voltammetry features for cross-dataset lithium-ion battery state of health prediction


Bağdat Ö. K., Tetik T., Avsar R., KONAR M.

Measurement: Journal of the International Measurement Confederation, cilt.292, 2026 (SCI-Expanded, Scopus)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 292
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1016/j.measurement.2026.123466
  • Dergi Adı: Measurement: Journal of the International Measurement Confederation
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO), Materials Science & Engineering Collection (ProQuest), Technology Collection (ProQuest)
  • Anahtar Kelimeler: Battery health estimation, Cross-dataset transfer, Differential thermal voltammetry, Feature engineering, Lithium-ion battery, Machine learning
  • Erciyes Üniversitesi Adresli: Evet

Özet

Lithium-ion battery state of health (SoH) models trained on laboratory data routinely fail under noisier field conditions, and the differential thermal voltammetry (DTV) literature lacks a systematic classification of which features survive measurement uncertainty versus cross-dataset transfer. Six DTV and two incremental capacity analysis (ICA) features are evaluated on the Oxford dataset under a measurement-chain perturbation budget spanning the temperature channel (white Gaussian, AR(1), 1/f, slow drift, and fixed-span ADC quantization) and the voltage channel (Gaussian noise of 1–10 mV, 12–16-bit quantization), and applied cross-dataset to NASA PCoE. An anti-diagonal 2 × 2 robustness taxonomy emerges: the voltage-position and ICA descriptors change by at most 3.7% under every perturbation tested, but their correlation with SoH does not transfer across cell designs, whereas the thermal-amplitude descriptors transfer within the LiCoO2/graphite chemistry family but degrade under temperature noise. No descriptor combines both properties. A hybrid fusion of DTV, ICA, and simple statistics reduces Oxford leave-one-cell-out root-mean-square error (RMSE) by 36% (paired Wilcoxon p<0.05 for gradient boosting and random forest) and is the only configuration with positive temporal R2. Cross-dataset transfer without adaptation fails: feature fusion reduces the Oxford-to-NASA error by 26%, yet no transferred model outperforms a constant predictor at the test-set mean (9.5% RMSE), and the lowest recorded transfer error (10.7%) is produced by a kernel model whose output degenerates to a constant. DTV features should therefore be selected by deployment scenario, and cross-cell-design deployment requires recalibration or domain adaptation.