Bayesian filtering techniques for state of charge and state of health estimation in lithium-ion batteries for electric vehicles: Methods, challenges, and future directions

Lithium-ion batteries (LIBs) play a pivotal role in the advancement of electric vehicles (EVs) because they offer excellent energy capacity, longevity, and versatility across applications. Monitoring state of charge (SOC) and state of health (SOH) precisely is essential for battery safety and performance, but these states cannot be measured directly due to the nonlinear and interconnected behavior of the batteries. While the conventional model-based, data-driven, and hybrid approaches enhance SOC estimation, they frequently suffer from high computational demands, instability, and limited adaptability. To handle system noise, nonlinear dynamics, and uncertainty, Bayesian filtering methods—including Kalman filters (KFs) and particle filters (PFs)—have been shown to be both effective and efficient. This review provides a comprehensive and structured overview of Bayesian filtering techniques for SOC estimation and SOC–SOH estimation, encompassing both unified and multi-timescale estimation frameworks with offline-initialized and/or online-updated model parameters. This distinction reflects that inaccurate or time-varying parameters can introduce persistent SOC bias, while SOH degradation fundamentally alters battery dynamics during long-term operation. From a system-theoretic perspective, SOC, model parameters, and SOH evolve on distinct time scales—SOC varying rapidly, parameters drifting gradually, and SOH changing slowly due to cumulative aging—which motivates their separate and joint treatment within Bayesian filtering frameworks. Particular emphasis is placed on data usage characteristics, computational complexity, and engineering feasibility for practical battery management system implementation. Compared with existing surveys, this work introduces a unified classification framework that enables systematic comparison of traditional Bayesian filters, PFs employing KF-based proposal densities, machine learning (ML)-assisted Bayesian methods, and hybrid KF–PF approaches in terms of estimation accuracy, robustness, interpretability, and real-time applicability. Representative experimental results reported in the literature are consolidated to highlight the performance improvements of hybrid and ML-assisted Bayesian filtering methods under noise and aging effects, while explicitly revealing the associated computational and real-time implementation trade-offs. Overall, this review is distinguished by a unified organizational framework and an engineering-oriented perspective that directly relates estimation performance to computational complexity and real-time BMS constraints.