A unified deep learning optimization paradigm for Lithium-ion battery state-of-health estimation

State-of-Health (SOH) is critical to managing the lifespan of the battery energy storage system. For the data-driven-based method, explicit features have many benefits, but they cannot be constructed automatically. To maximize the use of explicit features and automate its construction and other configuration processes, this paper design a unified optimization paradigm for synergizing the three fundamental procedures in the data-driven model, that is, feature extraction, importance assignment, and model parameterization. An Evolutionary Multi-Objective (EMO) method is used to synchronously find a series of non-dominated solutions for the best combination of features, attention layer, and hyper-parameters of the network, which can enable a flexible SOH estimation in various operation conditions. A piecewise aggregate approximation is designed to compress the partial voltage curve while keeping its tendency characteristics. A Long Short-Term Memory (LSTM) is used to establish the data-driven model, especially an attention layer is added to finalize the task of feature selection. Experimental results prove the proposed model can achieve accurate SOH estimation and also provide flexible solutions for different scenarios. And results with multi-battery validation and transfer learning demonstrate that the proposed method not only has a high generalization ability but also easily transfers to a new task.