Robust Data-Driven State of Health Estimation of Lithium-Ion Batteries Based on Reconstructed Signals

The state of health (SoH) of lithium-ion batteries is critical for diagnosing the actual capacity of the battery. Data-driven methods have achieved impressive accuracy, but their sensitivity to sensor noise, missing samples, and outliers remains a limitation for their deployment. This paper proposes a robust, purely data-driven SoH estimation methodology that addresses these challenges. Our method uses a proposed non-iterative closed-form signal reconstruction derived from a modified Tikhonov regularization. Five new features were extracted from reconstructed voltage and temperature discharge profiles. Finally, a Huber regression model is trained using these features for SoH estimation. Six ageing scenarios built from the public NASA and Sandia National Laboratories datasets, under severe Gaussian noise conditions (10 dB SNR), were employed to validate our proposed approach. In noisy environments and with limited training data, our proposed approach maintains a competitive accuracy across all scenarios, achieving low error metrics, with an RMSE on the order of 10 − 4 , an MAE on the order of 10 − 2 , and a MAPE below 1%. It outperforms state-of-the-art deep neural networks, direct-feature Huber models, and hybrid physics/data-driven models. In this work, we demonstrate that robustness in SoH estimation for lithium-ion batteries is influenced by the choice of machine learning architecture, loss function, feature selection, and signal reconstruction technique. In addition, we found that tracking the time to minimum discharge voltage and the time to maximum discharge temperature can be used as effective features to estimate SoH in data-driven models, as they are directly correlated with capacity loss and a decrease in power output.