Transfer Learning from Synthetic Data for SOH Estimation

Public data on battery cell and battery pack aging, especially accessible from outside the automotive industry, is scarce. Scaling up from cell to pack level is not straightforward, as it requires an understanding of the underlying inhomogeneities and degradation patterns. Within the automotive industry, comprehensive time-series data on battery pack or battery electric vehicle aging is also limited due to high costs and data privacy concerns related to customer fleet data. Moreover, it is nearly impossible to capture the vast array of possible aging trajectories in controlled experiments. Such datasets, however, are essential for diagnosis and forecast of battery pack aging in the field, as well as for developing data-driven methods for state estimation. In this work, we introduce a fast, publicly available, battery pack aging data simulation toolbox based on pristine half-cell potential measurements at various C-rates. This toolbox enables the generation of constant-current charging events at different C-rates and can simulate any conceivable aging path. The modular framework allows users to adjust individual cell parameters, including state of charge (SOC), state of health (SOH), and degradation modes. Utilizing pristine half-cell potential measurements from a modern automotive lithium-ion battery, we create an initial cell model. This model, founded on the ‘alawa-toolbox [1], is validated with experimental data from a laboratory dataset. As Figure 1 illustrates, we construct a battery pack model by serially and parallelly connecting multiple cell models, incorporating intermediate resistances. This pack model reflects the configuration of a physical battery pack and permits modifications to SOC, SOH, degradation modes, and lead resistances. We collect time-series data from partial charging events of development vehicles at various aging states and use this data to calibrate our model by minimizing the discrepancy between simulated and measured voltage curves. Our model allows users to simulate and analyze the evolution of battery pack asymmetries, creating a digital twin to monitor aging impacts and identify the "weakest" cell for early damage detection and intervention. Giving every user the opportunity to generate big data from our toolbox expedites the development of data-driven SOH estimation or open-circuit voltage (OCV) reconstruction models. Future research will explore the use of synthetic data to develop state estimation algorithms applicable to various chemistries and configurations, including the analysis and validation against real customer fleet data, their battery pack aging and asymmetry patterns. Our work represents a significant advancement in facilitating access to battery and battery electric vehicle aging data, thereby reducing barriers in this research field. We anticipate that our contribution will hasten the development of data-driven methods at the battery pack level. [1] M. Dubarry, C. Truchot, B. Y. Liaw, Synthesize battery degradation modes via a diagnostic and prognostic model, Journal of Power Sources 219 (2012) 204-216, https://doi.org/10.1016/j.jpowsour.2012.07.016. Figure 1