Estimating Lithium-Ion Battery Health Using Hybrid Attention Networks and Multisource Data

The accurate state-of-health (SOH) estimation for lithium-ion batteries (LIBs) is crucial for operational stability, longevity, and timely replacement in electric vehicles (EVs). Using fragmented data to extract aging features has become a popular research method for estimating SOH in a data-driven approach today. However, the existing fragmented aging information often suppresses effective aging features due to the presence of explicit noise from missing data. The fusion of new aging information is usually required to alleviate this deficiency. Additionally, due to advancements in sensor technology, an increasing number of high-precision, low-cost sensors are being introduced to the automotive manufacturing market, making the multidimensional detection of EV aging information feasible. Therefore, this study proposes a hybrid attention estimation framework that fuses electrical and electrochemical impedance spectroscopy (EIS) aging information and comprises two main components. The first part simulates the actual charging and discharging conditions of EVs to obtain the electrical aging information at random state of charge (SOC) intervals during the charging phase and the EIS impedance aging information at random cycle times. The second part introduces local, global, and cross-attention mechanisms (CAMs) to construct a hybrid attention mechanism network (HAMN). This network enhances the expression of aging information by employing the local attention mechanism (LAM) to filter out noise. Next, the electrical and EIS aging information are fused using the CAM. Subsequently, this fused aging information is fully utilized by the global attention mechanism (GAM), in conjunction with a fully connected neural network (FCNN), to establish the mapping relationship between aging information and SOH. The authors used laboratory equipment with Hall current sensors and charge-discharge machines for collecting battery aging data, supplemented by NASA’s experimental data. Results show that the network excels in SOH fragment information estimation.