SOH estimation framework for batteriesconsidering label normalization and feature stability under real-world data

To accurately estimate the state of health (SOH) of electric vehicle battery packs under operating conditions, it is necessary to overcome the challenges posed by inaccurate SOH labels and unstable estimation features in Real-World data. To overcome these challenges, we propose GAM-FNN, a deep learning framework that incorporates an attention mechanism and a grouped learning approach to effectively utilize multi-source physical data for precise SOH estimation. A novel SOH labeling strategy based on the interval capacity during the charging process is introduced, and its accuracy is further improved through normalization of temperature and initial charging voltage. Considering the complex behaviors such as charging/discharging, vehicle states, and consistency, our model fully leverages and optimizes these physical characteristics, particularly through the introduction of current distribution during acceleration as a new feature to enhance SOH estimation. To address the instability and limitations of features under real-world operating conditions, we design a grouped learning mechanism based on the physical significance of features and employ a stability-weighted attention mechanism to improve the model's robustness. We evaluate our framework using a dataset constructed from four years of historical data of 860 EVs operating across a wide range of temperature regions. With SOH after three years of operation as the estimation target, the results demonstrate outstanding accuracy, achieving a mean absolute percentage error (MAPE) of just 2.337 % in estimating.