• 대한전기학회
Mobile QR Code QR CODE : The Transactions of the Korean Institute of Electrical Engineers
  • COPE
  • kcse
  • 한국과학기술단체총연합회
  • 한국학술지인용색인
  • Scopus
  • crossref
  • orcid
Title ICA Based Transfer Learning Strategies to Predict the Life of Aging Batteries in Various Operating Environments
Authors 하태빈(Tae-bin Ha) ; 이상력(Sang-ryuk Lee) ; 김태윤(Tae-yoon Kim) ; 송민우(Min-woo Song) ; 김종훈(Jing-hoon Kim)
DOI https://doi.org/10.5370/KIEE.2026.75.5.1077
Page pp.1077-1085
Keywords Lithium-ion battery; Sequence-to-Sequence; Fine-tuning; Transfer learning; Incremental Capacity Analysis(ICA)
Abstract Lithium-ion batteries exhibit different degradation pathways under diverse operating environments, and in real-world applications, reliable life prediction is often required using only early-stage operating data before sufficient long-term history is accumulated; considering the time and cost of data collection, transfer learning becomes essential. However, when operating conditions change, data distributions readily shift, and prediction performance can deteriorate if a model trained under a single condition is directly applied or if the transfer intensity is determined solely based on time-series linearity or distribution-level similarity. To mitigate both domain discrepancy and limited early-data constraints, this study proposes an ICA-guided transfer learning strategy that determines the transfer intensity using only the first 200 cycles. First, eight aging-profile datasets were obtained using INR21700-33J cells, and a Seq2Seq?LSTM pre-trained model was developed using capacity- and voltage-based health indicators augmented with rate-of-change and exponential moving average features. During transfer, initial similarity was assessed via Pearson correlation of the 200-cycle capacity time series; however, we found that time-series similarity alone can lead to performance degradation under certain conditions. To address this limitation, we additionally incorporated similarities of Peak 1 and Peak 2 extracted from incremental capacity (IC) curves. Specifically, when IC-peak correlation was high, conservative transfer focusing on the output layer was applied, whereas low IC-peak correlation triggered stronger transfer by unfreezing the decoder along with the output layer. The results demonstrate that the proposed ICA-peak-based decision rule consistently secures high explanatory power across all target sets with stable prediction performance.