A methodology for lithium-ion battery state of health estimation using random constraints of state of charge

Accurate estimation of the state of health (SOH) of lithium-ion batteries is critical for reliable operation and lifetime management. Conventional SOH estimation methods typically require long testing durations over full voltage or state-of-charge (SOC) ranges, while approaches relying on limited low-voltage information often suffer from reduced accuracy. To address these challenges, this paper proposes an excitation-aware SOH estimation framework based on a BiLSTM–multi-head self-attention network. Aging data collected under multiple excitation conditions are utilized, and feature parameters extracted from constrained voltage intervals are jointly analyzed to capture localized degradation characteristics. By modeling the nonlinear relationship between voltage-interval-based features and SOH, the proposed method achieves high estimation accuracy within five consecutive 0.1 V voltage intervals, with a maximum error of 1.61%. Furthermore, validation on lithium-ion cells with different capacities and SOH levels demonstrates strong generalization capability, with estimation errors remaining within 1.75%. The proposed approach enables accurate SOH estimation under random SOC conditions, providing an effective solution for rapid and practical battery health assessment. • Construction of a multidimensional feature fusion framework for SOH estimation under random SOC conditions. • Development of a BiLSTM-Transformer hybrid deep learning model for high-accuracy SOH estimation. • Cross-cell and multi-rate validation with quantitative error evaluation.