A data-efficient method for lithium-ion battery state-of-health estimation based on real-time frequent itemset image encoding

Next-generation intelligent battery management systems (BMS) require accurate real-time estimation of battery state of health (SOH). However, existing studies often underestimate challenges arising from large volumes of online data with varying quality, as well as the resulting pressures on data storage, transmission, and computation. This paper proposes a lossy counting-based gated dual-attention Transformer (LC-GDAT) framework that substantially reduces historical data storage needs while maintaining high accuracy in SOH estimation. To overcome errors due to information loss from data compression, two critical modules are introduced. The first is the parallel temporal-spatial lossy counting feature extraction module (PTS-LC). It uses frequent-item extraction to identify important voltage and charging capacity patterns during battery operation. This significantly reduces storage demands and effectively transforms frequent items into two-dimensional features. The second module is the gated dual attention Transformer (GDAT). It uses a dual-branch structure to adaptively explore battery degradation characteristics from positional and channel dimensions. A gating mechanism is introduced to enhance interaction between these dimensions. The performance of LC-GDAT is comprehensively evaluated using data from 124 batteries under laboratory conditions, as well as real-world data from 20 electric vehicles collected over approximately 29 months. The experimental results show that LC-GDAT achieves the lowest SOH estimation errors of 0.46 % under laboratory conditions and 2.23 % under real-world conditions.