Diagnosing failures in lithium-ion batteries with Machine Learning techniques

The industry of electric cars has been rising significantly in the last few years. This significant growth is impulsed by lithium-ion batteries (LIBs). LIBs are a stable technology that is the largest applied for storing energy systems in several sectors like electric mobility, laptops, and cell phones. However, LIBs are still sensible and must operate under controlled conditions to avoid failures and anomalies that can cause security and performance issues. Therefore, it is crucial to identify the possible failures in the initial stage. In order to help the precocious failure identification, some data-driven have been impulsed by Machine Learning (ML) techniques to diagnose failures as soon as possible. Following this direction, this work presents the application of six classification ML models to identify failures in LIBs. In order to generate the dataset, the main parameters of eight lithium-ion 18650 cells were collected under four different cases of abuse: overcharging (OC), over-discharging (OD), external short-circuit (ESC) and overheating (OH). The confusion matrix and Sensitivity were used to evaluate and compare the models. The work shows that, on average, the models could detect a failure with the Sensitivity greater than 94% for mechanical, electrical, and thermal failures. Furthermore, the RF model had the better performance of all, with the Sensitivity greater than 97%. In conclusion, according to the excellent results, this approach can be used in LIBs to detect failures in the initial stage. Due to the excellent results, the evolution of this study is embedded this study in the vehicle to classify abuses online to reduce the chances of security and performance issues.

Automated summary

The rapid growth of the electric car industry in recent years is driven by the use of stable lithium-ion batteries (LIBs). Although LIBs are widely used in energy storage systems for electric mobility, laptops, and cell phones, they are still prone to failures that can pose security and performance issues. To identify failures at an early stage, this study applies six classification machine learning models to LIBs, achieving a sensitivity rate of over 94% for mechanical, electrical, and thermal failures. The random forest model performs the best with a sensitivity rate over 97%. This approach can effectively detect failures in LIBs and has the potential to be integrated into vehicles for online abuse classification, reducing security and performance risks.