1 December 2019 in Tetrahedron Letters
22 August 2025 in IEEE Transactions on Instrumentation and Measurement
Accurate state of health (SOH) prediction and data imputation are critical for advanced battery management systems, but remain challenging due to complex degradation patterns, diverse operating conditions, and varying battery materials. Although transfer learning methods have demonstrated potential in dealing with battery SOH prediction with various conditions, their need for fine-tuning limits real-world applications. To address these challenges, the feature enhanced structured state space diffusion model (FESD) is proposed for SOH prediction and imputation across diverse working conditions. In FESD, the framework first extracts degradation features from raw battery cycle data through feature engineering. These features, along with SOH data, are then processed by a Transformer-based architecture to capture intricate degradation patterns and long-range dependencies. Finally, the processed features are fed into a diffusion model built upon structured state-space equations, which effectively captures temporal dynamics and allows for simultaneous SOH prediction and imputation. Evaluations on the MIT, CALCE, NASA, and HNEI datasets demonstrate FESD’s superior performance in SOH prediction and imputation. Notably, when applied to a new dataset without fine-tuning, FESD achieves comparable performance to state-of-the-art transfer learning methods, illustrating its strong generalization capabilities.
13 April 2026 in IEEE Transactions on Transportation Electrification
Accurate estimation of the state of health (SOH) and remaining useful life (RUL) of second-life lithium-ion batteries (SL-LIBs) retired from electric vehicles is critical for their safe reuse in energy storage systems. Existing methods often face challenges such as limited labeled data and poor generalization across diverse working conditions. To address these limitations, we propose a co-estimation framework for SL-LIBs that integrates self-supervised auxiliary learning, gradient adversarial training, and uncertainty calibration within a multi-task learning architecture. The framework introduces the discharge capacity change rate as an auxiliary label to enrich degradation information, while gradient adversarial training mitigates gradient interference between tasks. Additionally, predictive uncertainty is quantified using Monte Carlo Dropout and further calibrated through isotonic regression to enhance reliability. Extensive experiments on a self-tested dataset of 84 SL-LIBs demonstrate that the proposed method achieves a mean absolute error of 0.0089 for SOH and 67 cycles for RUL, reducing the best published baseline by 41% and 39%, respectively.
25 May 2026 in Renewable and Sustainable Energy Reviews