A Transfer Gated Recurrent Unit Model for Battery Health Monitoring Under Cross-Domain Conditions

Battery health monitoring is crucial for ensuring maintenance reliability and safety, with broad applications in various industrial systems. Data-driven state-of-health (SOH) estimation methods can perform excellently using massive historical data without prior physical knowledge. In the battery aging process, this article extracts the essential health indicators to construct the model and adopts a suitable data preprocessing method to reduce the influence of noise. There is still a limitation for data-driven methods regarding data quality, and the domain shift caused by different operating conditions will damage estimation performance. To tackle the issue, this article proposes a transfer feature transformation (TFT) based on mean maximum discrepancy (MMD) to get a new feature space to get domain-invariant features with sufficient information. We propose a multilayer gated recurrent unit (GRU) neural network with fully connection (FC) layers to uncover hidden relationships in degraded data. Fusing TFT and GRU-FCs methods can realize an end-to-end SOH estimation under cross-domain conditions. We apply real-world public Center of Advanced Life Cycle Engineering (CALCE) and NASA battery datasets and lab testing-based accelerated aging data to validate the performance of algorithms. By comparison, our proposed TFT-GRU-FCs method can get the best performance of all algorithms, and we have also given sufficient results as support.