State of health (SoH) estimation is crucial to verify safe, efficient and reliable performance of lithium-ion batteries (LIBs) in electric vehicles and energy storage systems. However, the aging mechanism of the battery decreases stochastically over several instances. To address the SoH estimation, a fusion-based framework is proposed by combining complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), multi-head attention mechanism bi-directional long short-term memory (MHAM-BiLSTM) and random forest (RF). Initially, the health indicators (HIs) are selected by evaluating the strong Pearson correlation coefficients, based on increasing or decreasing voltage segments within the constant-current/constant-voltage charge–discharge cycles. These HI data are pre-processed by CEEMDAN, which decomposes signals into intrinsic mode functions (IMFs) and residual (RES) components to enhance the characteristics of the signal. The importance of IMFs is evaluated using the variance contribution rate (VCR). Then, the IMFs are processed through MHAM-BiLSTM, and their extracted characteristics are weighted by VCR to estimate SoH. The RES component is fed into the RF model. The outcomes of both techniques are fused to extract a more accurate and robust SoH prediction via the proposed CEEMDAN-VCR-MHAM-BiLSTM-RF model. The proposed model validates notable improvements in accuracy and computational resources of the estimated SoH, which validates effectiveness.