Accurate state of health estimation based on the multi-feature extracted and improved MHA assisted CNN-BiGRU model for lithium-ion batteries

Lithium-ion batteries (LIBs) inevitably experience aging degradation during operational lifespan, making accurate state of health (SOH) estimation essential for ensuring battery safety and reliability. Existing single-attribute network often exhibits limited accuracy, robustness, and generalizability capabilities. Meanwhile, the sensitivity of traditional incremental capacity (IC) curves to noise and insufficient capture of aging features further impacts estimation precision. To address these gaps, a hybrid framework integrating noise-resilient feature engineering is proposed for estimating SOH accurately. First, the smoothed IC profiles generate from the reconstructed charging voltage-capacity curves, effectively alleviate the noise compare with conventional one. Second, a comprehensive set of aging features is extracted from both voltage and enhanced IC curves, capturing multidimensional degradation patterns overlooked by single-feature approaches. Finally, a novel network architecture combining multi-head attention (MHA), convolutional neural networks (CNN), and bidirectional gated recurrent units (BiGRU) is developed. The MHA mechanism dynamically weights critical temporal features to mitigate overfitting risks inherent in existing recurrent models, while CNN and BiGRU synergistically learn local degradation trends and long-term aging dependencies. Moreover, the catch fish optimization algorithm (CFOA) automatically adapts hyperparameters, eliminating manual tuning biases. The simulation results demonstrate that the proposed method achieves excellent accuracy, robustness, and generalizability across two distinct battery datasets, diverse aging conditions, and different training strategies. Additionally, its superiority over state-of-the-art methods is validated through comprehensive comparative analyses.