battery health New machine learning chip could help prolong battery life in smart devices
Researchers at the University of Cambridge have devised a machine learning method that can predict battery health with 10x higher accuracy than the current industry standard.
This is potentially a big deal. If the research team’s method can be applied commercially, it could be a boon for the development of safer and more reliable batteries for a myriad of applications, including electric vehicles and consumer electronics like smartphones and wearables.
Working in collaboration with Newcastle University, the Cambridge-led research team designed the new monitoring method by sending electric pulses into batteries and then measuring the response. The measurements are then picked up and processed by a machine learning algorithm to predict the battery’s health and its useful lifespan (i.e., how long it will last for under normal use and conditions). The team’s method is completely non-invasive and can simply be added on to existing battery systems.
The problem with batteries
Over time, batteries fail. This is one of the biggest challenges facing modern electronics that demand more from their power sources, and one application where it is presenting a particularly damning challenge is that of electric vehicles. Over time, a complex network of subtle chemical processes cause batteries to degrade and while individually these processes do not have much of an impact on battery performance, collectively they shorten its performance and lifespan.
As such, being able to predict the state of a battery’s health and its remaining useful lifespan could be a game-changer for not only the EV industry but others too, such as consumer electronics where battery degradation leads to a requirement for more frequent charging—quite the annoyance to which we can all relate.
Current methods for predicting battery health rely on tracking the current and voltage during charging and discharging. However, this misses many important features that are vital indicators to its health and tracking these features and processes that are taking place within a battery requires new ways of probing them while they are in action, as well as algorithms that can detect subtle signals during charge cycles.
"Safety and reliability are the most important design criteria as we develop batteries that can pack a lot of energy in a small space," said Dr. Alpha Lee from Cambridge's Cavendish Laboratory, who co-led the research. "By improving the software that monitors charging and discharging and using data-driven software to control the charging process, I believe we can power a big improvement in battery performance."
The research team’s method
The team designed a way to monitor battery health by sending electrical pulses into it and measuring the response. Once a response has been received, a machine learning model is employed to look out for specific features in it that are indicative of battery aging. The team trained this machine learning model with over 20,000 experimental measurements, representing the largest dataset of its kind. Most importantly, the model can distinguish important signals from irrelevant fluff and noise, making it highly accurate.
The research team also demonstrated that this machine learning model can also give hints about the physical mechanism of degradation, such as which signals correlate most with aging. This allows the team to design focused experiments to probe why and how batteries degrade.
“Machine learning complements and augments physical understanding," said co-first author Dr. Yunwei Zhang, also from the Cavendish Laboratory. "The interpretable signals identified by our machine learning model are a starting point for future theoretical and experimental studies."
Now, the team is using their machine learning method to investigate and understand degradation in different battery chemistries.