Remaining Discharge Time Prognostics of Lithium-Ion Batteries Using Dirichlet Process Mixture Model and Particle Filtering Method

A new approach using Dirichlet process mixture model (DPMM) and particle filtering (PF) method to predict remaining discharge time (RDT) of ongoing discharge processes of lithium-ion batteries is proposed. Different voltage trajectory patterns are proposed to describe the discharge process at different periods of a battery's life. Each pattern is represented by the same empirical model based on the physical discharge behavior of lithium-ion batteries, with different parameters to distinguish itself. A DPMM is developed to automatically discover these voltage trajectory patterns from historical monitoring data, without specifying the number of patterns in advance. The trajectory parameters for each pattern can also be learned simultaneously, which are used as the initial parameters for online prognostics. The developed DPMM is able to discover new patterns as more trajectory data become available. During online prognostics, voltage trajectory pattern is constantly identified with new voltage data, then initial parameters for identified pattern and PF-based method are combined to predict the RDT of the ongoing discharge process. A case study demonstrating the proposed approach is presented. It also demonstrates that this approach improves accuracy of RDT prediction compared with benchmark PF-based method.