Battery Diagnostics & RUL

Upgrading from electrochemical models to machine learning for more dynamic analysis.                       

Share on linkedin
LinkedIn
Share on twitter
Twitter

Objectives

  • Demonstrate viability of using machine learning for electric vehicle battery diagnostics and remaining useful life (RUL) estimation
  • Exceed performance of standard electrochemical approach

Challenge

  • Minimal dataset
  • High variability in usage patterns

Results

  • Acerta created a machine learning model for survival analysis of electric vehicle batteries
  • The machine learning model performed better than the standard electrochemical approach, offering the potential for huge savings to OEMs through extended battery life

Background

The global electric vehicle (EV) industry is expanding rapidly, increasing the demand for more efficient battery technologies. While lithium-ion (Li-ion) batteries remain the industry standard, their use in EVs presents a novel set of risks for automotive OEMs, including fires due to thermal runaways, difficult-to-predict material aging, and the unanticipated loss of capacity over time as a result of variable driver and charging behaviours.

Given the significant cost of battery replacement and the requirement that manufacturers offer warranties of at least 8 years, there is a pressing need for more accurate estimates of remaining useful life (RUL) for EV batteries.

The Problem

An EV Li-ion battery reaches the end of its life when its capacity drops to 80%. At this point, the EV loses significant driving range. Traditionally, manufacturers use oversizing to compensate for the inevitable drop in capacity over time, increasing the size of battery packs but also adding additional cost and weight to the vehicle.

While OEMs can use electrochemical models of batteries to estimate RUL, these are inherently static, and thus unable to account for variabilities in a vehicle’s environment, as well as driving and charging behaviour. As a result, even the best electrochemical models will struggle to predict faults or generate RUL estimates that can be applied to every battery in a set with equal confidence.

Solution Process

Using a combination of open-source and proprietary data, Acerta was able to build up a data-set of battery tests involving both standard usage and accelerated aging scenarios. The tests also used different types of Li-ion batteries with varying numbers of cells. In some cases, the batteries were tested up to a certain point of capacity loss, while in others they were operated under normal conditions with consequently little to no degradation.

The resulting data encompassed a variety of crucial parameters for assessing battery performance, including voltage, current, temperature, capacity, and the time, date and location of operation. This was combined with a collaborative effort between Acerta and the University of Waterloo’s Chemical Engineering department. By building equivalent circuit models to EV batteries, the Waterloo engineers were able to simulate various behavioral profiles, such as charging and discharging, and changes in friction, heat, and vibration. The resulting electrochemical model gave Acerta a benchmark against which to test our machine learning models.

One of the immediate benefits of this approach was that it helped identify just how much variation there is in battery profiles, not only from charging and discharging patterns, but also aging, driving behavior, and seasonal changes in external factors, such as temperature and humidity. This is where machine learning thrives: responding dynamically to how a battery is being used over time.

Results

Acerta created a machine learning model for battery survival analysis based on the probability that a battery will hit its capacity threshold over the course of its lifetime. Despite the lack of available data and the high degree of variability in what was available, Acerta’s machine learning model demonstrated levels of accuracy and precision as good or better than the electrochemical model.

With EV batteries each costing $5,000 – $15,000 to replace, a user notification system that uses machine learning to suggest changes in charging or driving behaviour to prolong battery life could offer significant savings to consumers and automakers both. As of 2019, there were 7.9 million EVs worldwide: extending the battery life of 1% of those vehicles by one year through AI recommendations, would result in savings of $400M – $1.2B annually.

For more information...

Review other case studies or visit our blog.

How can we Help?

Contact Info

Subscribe to our Monthly Newsletter for more Automotive Intelligence!

This website uses cookies.

Welcome Back!

Stay on the cutting edge of automotive Intelligence

Subscribe to Our Monthly Newsletter for Industry Trends, Best Data Practices & Industrial Machine Learning Applications