A battery SOH estimation method based on PI-TFT-iDOA driven Li-battery discharge state features

Accurately estimating the health status of lithium batteries is crucial for their safe use. The existing methods, including those based on a single model or simple fusion ones, often struggle to fully exploit the deep coupling characteristics between multimodal data. For this purpose, this paper proposes a fusion-based SOH estimation framework (PI-TFT-iDOA) that integrates temporal model and physical constraints through a polarity-cross-attention mechanism. Firstly, extract discharging time, degeneration of energy and capacity increment (IC) area as health features, and use maximum correlation minimum redundancy (mRMR) to analyze the relationship between indicators and capacity decay. Through analyzing population parameter trends with finite gradient differences and Hessian matrices, the exploration phase of Dream Optimization Algorithm(DOA) is partitioned into deep, shallow, and REM stages. This staging strategy effectively enhances the hyperparameter optimization performance for state-of-health (SOH) estimation by executing optimal solution search within dynamically adjusted trust regions. Results of SOH estimation for 12 different batteries show the proposed method achieves high accuracy with an average MAE of 0.5 % and RMSE of 0.94 %, which is notably lower than the RMSE of 1.65 % reported by CNN-LSTM across varying training data volumes.