A hybrid variational mode decomposition and optimized long short-term memory framework for Lithium-Ion battery state-of-health estimation with multidimensional features

Accurate state-of-health (SOH) estimation is a core requirement for reliable lithium-ion battery management systems. This paper proposes a hybrid framework integrating multidimensional feature engineering, adaptive signal decomposition, and automatic hyperparameter optimization. First, eight health features covering current, voltage, temperature, and incremental capacity domains are extracted, and four highly SOH-correlated features are screened via training-set-restricted Pearson analysis to avoid information leakage. Variational mode decomposition (VMD) is then applied to decompose feature sequences into multi-scale intrinsic mode functions to separate long-term degradation trends from measurement noise, followed by principal component analysis (PCA) to eliminate inter-component redundancy and retain dominant information. A dung beetle optimizer (DBO) is further introduced to automatically tune long short-term memory (LSTM) hyperparameters including learning rate, the number of neurons in each layer, and dropout rate, thereby overcoming performance fluctuations caused by manual tuning. Validation on NASA PCoE and CALCE datasets shows that the proposed VMD-PCA-DBO-LSTM framework reduces RMSE by up to 30.5% compared with the benchmark model, achieving minimum RMSE of 0.0029 and 0.0159, with maximum R 2 of 0.9927 and 0.9953, respectively. Notably, the model maintains stable performance even with only 50% training data, confirming its robustness for practical BMS deployment under limited-data scenarios.