Lithium-ion batteries have become the backbone of today’s clean energy revolution, powering everything from electric vehicles and consumer electronics to renewable energy storage. As their use grows worldwide, a new challenge has emerged: how to measure and manage the health of these batteries over time. This is where the concept of State of Health (SoH) becomes critical. SoH represents how much a battery has aged compared to its original state and reflects its ability to safely deliver energy. In simple terms, SoH tells us how “healthy” a battery really is.
Accurate measurement of SoH is essential because it directly impacts safety, performance, and cost. Without reliable SoH monitoring, batteries may be retired too early—wasting valuable resources—or too late, creating risks of failure and safety hazards. For electric vehicles and large-scale energy storage, real-time SoH monitoring also enables smarter battery management, ensuring better charging, discharging, and thermal regulation. This not only extends battery life but also supports safer and more sustainable use.
In electric vehicles, lithium-ion batteries are central to their performance, but even after reaching the end of their first life, they still hold significant capacity. These second-life batteries can be reused in less demanding applications such as stationary storage for solar and wind energy, backup systems for homes, or local microgrids. Reusing EV batteries in this way helps reduce environmental waste and provides cost-effective solutions for energy storage in different sectors.
Our research focuses on developing practical and accurate methods for SoH estimation, especially for these second-life batteries. Traditional laboratory approaches often struggle under real-world conditions, where noise and variability make predictions less reliable. By applying advanced machine learning techniques with electrochemical impedance spectroscopy (EIS) data, we aim to bridge this gap—offering solutions that are not only accurate but also robust, scalable, and ready for deployment in real-world energy systems.
In this study, we focused on estimating the State of Health (SoH) of lithium-ion batteries using advanced data-driven methods. To achieve this, we relied on Electrochemical Impedance Spectroscopy (EIS) tests, a powerful diagnostic technique widely used to study battery behavior.
What is EIS?
EIS works by applying a small alternating current (AC) signal to the battery over a wide range of frequencies and then measuring how the battery responds. This test provides a frequency-domain fingerprint of the battery, revealing valuable information about internal processes such as ion transport, charge transfer resistance, and overall degradation. Because every “aged” battery responds differently, EIS is a non-invasive way to detect hidden health conditions without dismantling the cell. we used seven key features extracted from Electrochemical Impedance Spectroscopy (EIS) tests as the input to our machine learning models. These features capture the essential characteristics of battery behavior across different frequencies and provide a reliable basis for accurate State of Health (SoH) prediction.
Applying Machine Learning
We trained and tested several machine learning algorithms using seven EIS-derived features to predict SoH. The Random Forest Regressor achieved the best results, with a prediction error as low as 0.08 RMSE, showing strong accuracy and robustness even under noisy, real-world-like conditions.
The two scatter plots below compare predicted and actual SoH values. Their close alignment highlights the reliability of our method and confirms that combining EIS features with advanced algorithms provides a practical framework for battery health monitoring.
To ensure our model could work reliably in real-world conditions, we tested it against different types of artificial noise that mimic common errors in battery management systems (BMS). Three scenarios were considered: Gaussian noise, representing random electrical interference; Uniform noise, simulating evenly spread measurement errors; and Bias shift, modeling small calibration drifts in sensors. These tests allowed us to evaluate how well the Random Forest model performs when faced with imperfect and noisy data.
Unlike many studies that only add noise during testing, we introduced these patterns both in training and testing. This “noise-augmented” strategy, combined with a simple outlier-clipping step, helped the model learn how to deal with uncertainty while maintaining strong accuracy. As shown in the results, the prediction error (RMSE) stayed below 1.4% in all cases, proving that the approach is robust and practical. By training the model to handle noise, we make it much more dependable for real-world battery systems, where sensor measurements are rarely perfect.
Our study demonstrates that combining Electrochemical Impedance Spectroscopy (EIS) with advanced machine learning models provides a powerful framework for reliable battery health estimation. By extracting seven key features from the EIS spectrum and applying them to algorithms such as Random Forest, we achieved highly accurate predictions of the State of Health (SoH). The results show that data-driven approaches can bridge the gap between laboratory analysis and real-world battery management, offering fast, non-invasive, and scalable diagnostics.
Beyond accuracy, the integration of noise-augmented training proved essential in making the model robust under realistic operating conditions. Even with different types of simulated sensor errors, the model consistently maintained low prediction errors, highlighting its readiness for deployment in practical applications. These findings underline the potential of combining physics-based insights with machine learning to extend the lifetime of lithium-ion batteries, enable safe second-life use, and ultimately contribute to building a cleaner and more sustainable energy future.
Nika Rahmani- Researcher at AICER Lab