Quantifying the Information Ceiling of Classical Machine Learning for BMS-Constrained Battery State-of-Health Estimation

Abstract

In this paper, the predictive ceiling of the classical machine learning-based lithium-ion battery state-of-health (SOH) estimation is estimated using 4 signal constraints. Current monitoring is also available for any production battery management system. Also available are the state of charge, terminal voltage, load current, current monitoring, and pack temperature. Random forest (RF), support vector support vector regression (SVR), and K-nearest neighbours (KNN) were used, and in this test, XGBoost was compared with 1,200 charge/discharge records. Laboratory cycled NMC lithium-ion pouch cells. All the R2 values that were lower than the baseline was returned by the four algorithms on the held-out test. set (SVR: −0.026; RF: −0.076; KNN: −0.119; XGBoost: This test was cross-validated five times, and the results were stable: −0.031. These results can be considered information theoretic; the four-dimensional instantaneous feature space is called diagnostic and is not enough to trace back the path-dependent history encoded in the SOH. The maximum value of the Pearson correlation coefficient (ρ) was 0.41. A flat, for SOC) and RF importance scores validate a flat, information-limited feature distribution. In this paper, the major constraint for classical shows is that they have information available. This is an ML-based SOH estimation under the BMS constraint. conditions. The BMS-deployable 8-layer modular design architecture is introduced, and a baseline is provided, which can be reproduced. Left for further research on feature enrichment.