PHYSICS-INFORMED FEATURE ENGINEERING AND DEEP EMBEDDED CLUSTERING FRAMEWORK FOR LITHIUM-ION BATTERY STATE-OF-HEALTH ANALYSIS

Abstract

Lithium-ion batteries serve as the cornerstone of modern portable electronics, electric vehicles, and large-scale energy storage systems. However, their long-term reliability is constrained by progressive degradation mechanisms that affect capacity, internal resistance, and overall performance. Accurate prediction of battery State-of-Health (SoH) has there fore become essential for enhancing battery safety, optimizing maintenance cycles, and improving energy-management strategies. While significant progress has been made in machine learning–based SoH estimation, current methods typically rely on pre-selected or domain-specific features, leaving a substantial gap in the exploration of generalizable feature engineering frameworks. Motivated by this limitation, the present work aims to sys tematically investigate new mathematical and data-driven feature constructs derived from real-world battery cycling datasets. This report proposes astructured methodology for extracting, evaluating, andselectinghigh value features from raw operational measurements such as voltage, current, temperature, charge/discharge capacity, and cycle count. The focus is on building a standardized feature engineering pipeline that incorporates mathematical transformations, degradation-sensitive indicators, and time-series based statistical descriptors. Additionally, the study explores the correlation between temperature dynamics, voltage relaxation behaviour, differential capacity curves, and cycle-induced degradation trends. The derived features are assessed for their predictive relevance using machine learning models, with emphasis on model interpretability and robustness. The proposed experimental design aims to bridge the gap between electrochemical un derstanding and data-driven modelling by formulating features rooted in mathematical relationships such as rate of change, gradient estimators, integral measures, and non-linear transformations. The ultimate goal is to identify a compact yet informative subset of features that can generalize across different battery types and cycling conditions. Thiswork laysthefoundationforfuturedevelopmentofacomprehensivefeature-engineering framework for SoH prediction. The outcomes are expected to support enhanced diagnostic models, reduce dependence on handcrafted domain features, and contribute toward safer, more efficient battery-management systems.