Identifying degradation patterns of lithium ion batteries from impedance spectroscopy using machine learning

Lithium-ion batteries power many crucial modern technologies, but their unpredictable degradation hinders advancement. Accurate prediction of state of health (SoH) and remaining useful life (RUL) is vital for managing battery replacement and recycling. Traditional approaches model microscopic degradation mechanisms (solid-electrolyte interphase growth, lithium plating, active material loss), but this is unscalable. Data-driven approaches offer an alternative, using real-time, non-invasive measurements and machine learning to link measurements to battery health without modeling physical mechanisms. However, defining informative inputs and building robust statistical models remain challenges. Charging/discharging curves are commonly used inputs, but electrochemical impedance spectroscopy (EIS), which provides rich information on material properties and electrochemical reactions, is underutilized. EIS spectra are high-dimensional, making feature selection difficult. Existing approaches simplify the spectrum via equivalent circuit models (often non-unique) or by focusing on handpicked frequencies. This paper proposes using Gaussian process regression (GPR) with the entire EIS spectrum as input, letting the model select relevant features. This approach is tested using the largest known dataset of EIS measurements (over 20,000 spectra) from commercial Li-ion batteries under varying conditions.

Existing literature focuses on two main approaches for battery forecasting: microscopic mechanism modeling and data-driven methods. Microscopic modeling attempts to capture the physical and chemical processes leading to battery degradation, such as SEI layer growth, lithium plating, and active material loss. While providing valuable insights, this approach is computationally expensive and often struggles with the complexity of real-world battery systems. Data-driven methods, on the other hand, aim to bypass the need for detailed physical models by leveraging machine learning to directly map measurable battery characteristics to SoH and RUL. Commonly used inputs for data-driven models are features derived from charging and discharging curves. While effective, these methods may not capture the full complexity of the degradation processes. Electrochemical Impedance Spectroscopy (EIS) is a powerful technique that provides a rich dataset reflecting various aspects of the battery's internal state. However, the high dimensionality of EIS data makes direct use with traditional machine learning methods challenging. Previous work has attempted to simplify EIS data by fitting to equivalent circuit models or selecting specific frequencies. This paper explores a different approach, utilizing the full EIS spectrum as input to a Gaussian Process Regression model.

The study used 12 commercial 45 mAh LiCoO2/graphite coin cells, cycled in climate chambers at 25°C, 35°C, and 45°C. Each cycle comprised a 1C-rate charge and a 2C-rate discharge. EIS measurements were taken at nine stages of each even-numbered cycle, covering a frequency range of 0.02 Hz–20 kHz. Capacity loss was determined after each odd-numbered cycle. The dataset comprised over 20,000 EIS spectra. Gaussian Process Regression (GPR) was employed for both capacity estimation and RUL prediction. The GPR model used the entire EIS spectrum as input, without feature engineering, and an ARD (Automatic Relevance Determination) kernel was implemented to automatically determine the relevance of different frequencies to degradation. The model was trained on a subset of cells and tested on others, with a training-test split that ensured that the cells in the training group and testing groups experienced the same charge/discharge rates. This allowed for a fair comparison between the model's predictions and actual measurements. To benchmark the method, the same GPR model was trained and tested using features extracted from the discharging curves instead of EIS spectra. The performance of the two models was then compared. In addition to the above methodology, another multi-temperature model was trained to assess the performance of the model without precise prior knowledge of the battery cycling temperature, with the assumption that the future operating temperature of a battery will remain similar to its past operating temperatures.

The GPR model accurately estimated battery capacity and predicted RUL using only the current cycle's EIS spectrum, even without knowing the past operating conditions. The ARD kernel revealed that low-frequency impedance features were most predictive of degradation. The model outperformed a similar model using features from discharging curves, indicating EIS offers richer information on battery health. The multi-temperature model demonstrated that accurate capacity estimation and RUL prediction are possible even without precise knowledge of the cycling temperature. The key frequencies identified by the ARD method remained consistent across different temperatures.

The results demonstrate the effectiveness of combining EIS measurements with GPR for accurate battery health estimation. The use of the entire EIS spectrum as input avoids the need for subjective feature selection and allows the model to automatically identify the most relevant features. The superior performance compared to models using discharging curve features highlights the richness of information contained in EIS data. The ability to accurately predict RUL without complete knowledge of past operating conditions is particularly significant for real-world applications. The findings suggest that EIS measurements, especially in the low-frequency region, are crucial indicators of battery degradation and that the proposed GPR method is robust to various operating temperatures. These results have implications for battery management systems, enabling accurate SoH and RUL prediction for improved battery life and safety.

This research successfully demonstrated the potential of using EIS coupled with GPR for accurate battery state of health and remaining useful life prediction. The method's robustness to variations in operating temperature and its superior performance compared to methods relying on discharging curves strongly suggest its applicability in real-world battery management systems. Future research could investigate the model's performance under more varied operating conditions (e.g., different charge/discharge rates) and explore strategies for reducing the number of frequencies needed for accurate predictions.

The study's limitations include the relatively small number of batteries used in the dataset (12 cells) and the constant charge/discharge rates used throughout the experiment. This may affect the generalization of the results to batteries operating under different conditions. Furthermore, the assumption that the operating temperature remains consistent, while simplifying the problem, is not always realistic in real-world scenarios. Further research is needed to investigate the method's performance under more diverse and realistic conditions.