iTransPASOH: A physics-aware iTransformer framework for lithium-ion battery state-of-health estimation
Year: 2026
Type: article
Abstract: • A dual-stage parallel data processing framework is proposed that incorporates prior physics principles of lithium-ion battery aging, statistical calculations, and time series settings, providing diverse learning representations beyond purely multivariate correlation-dependent feature extraction. • A novel tendency-alignment soft constraint approach is developed using iTransformer architecture that aligns general degradation trends between predicted and actual sequences, overcoming training divergence issues caused by rigid monotonicity enforcement during capacity recovery effects. • The first theoretical derivation connecting Calendar and Cycle Aging (CCA) modeling to Physics-Informed Neural Networks (PINN) is established, where a generalized PDE formulation is mathematically derived and integrated into the loss function for physics-informed neural network training. • Temporal pattern capture is integrated into dynamic PDE modeling through physics-guided optimization, where both multivariate outputs and temporal patterns are combined to ensure robust and generalized SOH degradation estimation with proven convergence stability. • Comprehensive experimental validation demonstrates an average MAPE improvement of 20.942% over state-of-the-art methods across four benchmark datasets, with statistical significance at the 95% confidence level, superior robustness under noise conditions (16.7% vs. 131.5% degradation), and practical inference times suitable for industrial deployment. The safety issue of lithium-ion batteries during the charging and discharging process is of great significance. Hence, state-of-health estimation, a key indicator in the battery management system, is crucial for battery safety control, degradation optimization, and precise remaining life prediction. Researchers have proposed several methods for measuring state-of-health, including direct measurement and soft sensor approaches, such as first-principle modeling and data-driven methods. However, these methods have significant limitations during applications. Hybrid methods, while leveraging advantages from both approaches, still face limitations including lack of flexibility and architectural design-level interpretability, over-dependency on data, and robustness issues. To solve these problems, we propose a trend-capture physics-aware optimization method together with iTransformer to capture the temporal patterns of SOH degradation. Then, to dynamically model the PDE that describes SOH degradation, we theoretically derive the CCA model into a generalized PDE form to provide interpretability and input the multivariate features (cross-sectional characteristics extracted from each charging cycle) and time-series features (temporal patterns across multiple cycles) respectively to enhance robustness. Next, we construct a PDE loss term using the numerical output from the PDE modeling to physically guide the optimization of the PINN. Finally, comprehensive experimental validation across four benchmark datasets demonstrates that iTransPASOH achieves an average MAPE improvement of 20.942% over state-of-the-art methods, with statistical significance confirmed at the 95% confidence level, while maintaining only 16.7% performance degradation under noise conditions and practical inference times of 6.4ms per sample for real-world deployment.
Source: Knowledge-Based Systems
Authors: Yiwei Zha, Feng Yan, Chunjie Yang, Guodong Mao
Institutions: Northeastern University, Zhejiang University
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,430
OpenAlex ID: W7135246208