Cambridge and Newcastle Researchers Use AI to Better Predict Battery Health

Dev Kapadia, ’23, Machine Learning, 4/18/20

 

Figure 1: The above distributions demonstrate the process of gaussian process regression (GPR), the artificial intelligence algorithm used by the researchers in the article to predict battery health. As the data points are fed into the algorithm, the “prior” graph is generated, representing the distribution of functions at different values of x. GPR is a Bayesian approach, so the identification of one observation depends on the previous. As data points are gathered, they are inputted into the algorithm for future use as the series of observation, producing the “posterior” graph. Lastly, once all observations are made, the “prediction” graph can be made to predict hypothetical observations at different values of x as if they are made in the same series and therefore dependent on previous observations in the series.5

Source: https://commons.wikimedia.org/wiki/File:Gaussian_Process_Regression.png

Though developed only about thirty years ago, the lithium ion battery has made several major contributions to society. Because of the stability and rechargeability of the battery, it is found in many of the consumer electronics that we consider everyday items, including our phones, cameras, and alarm systems. They are also being used now in products of the “green revolution,” like electric vehicles and solar panels.1

Despite the widespread use of the lithium ion battery, researchers still don’t know exactly what the true health of these batteries are during use. Over time, the health of a battery deteriorates due to a variety of chemical processes that reduce the capacitance and the subsequent charge the battery can deliver.2 If battery health were better tracked, it would not only mean more accurate recommendations for battery replacement, but could also improve battery recycling, as scientists would be able to better identify which batteries have sufficient charge to be reused.3

Currently, the most common approach for battery health estimation is to model the various microscopic chemical processes that degrade battery health. However, there are a couple of issues with this virtual modeling. For one, some of these microscopic processes that are yet to be modeled – so their contributions to battery health are not fully accounted for. Second, some of these processes are so complex that modeling is unfeasible. Besides modeling, another method is to simply measure the current and voltage of the battery, but this method ignores the complexities determining battery health and is therefore a far less accurate predictor of battery life.3

In response to the large amount of data required to analyze the complex system of battery degradation, researchers from Cambridge and Newcastle Universities have developed a data-driven approach to estimating battery health using artificial intelligence.4 The researchers created a procedure that involves sending electrical pulses at varying levels of frequency into the battery; the response is then an “impedance spectrum” which is essentially measurements of current at the given frequency. The researchers then process the response with an algorithm to determine properties of the battery that will help determine battery health. This algorithm, trained with a dataset of over twenty-thousand experimental measurements, can distinguish responses that indicate battery degradation from those that are normal or unrelated to battery performance.3

The measurement method is known as electrochemical impedance spectroscopy (EIS), which measures the resistance of the system through the response to the various intensities of electric pulses sent to the battery. The response gives a lot of information about the physical properties of the system, but common methods have had to disregard much of the output data because of the inability to determine whether a result actually has an effect on battery health.3 To rectify this problem, the research team used a process known as gaussian process regression (GPR), which makes predictions on the impact of the various outputs based on probabilities determined by the previous measurement and frequencies within the dataset.3, 5

Using the combination of these two methods – EIS and GPR – the team was able to obtain battery health estimations with ten times higher accuracy than other existing methods. Even better, the electrical signals the researchers found to indicate battery degradation will allow them to construct more specific experiments to determine exactly how and why the batteries age.4 The team plans to use machine learning to study the mechanics of battery charging to develop optimal charging protocols that minimize the aging of batteries.4

While battery performance estimation is improving, there are still areas for improvement. With artificial intelligence being one of the many centers of attention in modern computer science, it seems only natural that this process would be used in an application that requires extrapolating current knowledge to predict unknowns (i.e. battery health). The Cambridge and Newcastle team hopes that the scientific community will choose to build off of the starting point that their battery degradation studies have created.4

 

Bibliography:

[1] Liu, Zhao. (11 Oct. 2019). “The History of the Lithium-Ion Battery.” Accelerating Microscopy, www.thermofisher.com/blog/microscopy/the-history-of-the-lithium-ion-battery/.

[2] Herrman, J. (2017, November 14). Why Your Gadgets’ Batteries Degrade. Popular Mechanics. https://www.popularmechanics.com/technology/gadgets/how-to/a7432/why-your-gadgets-batteries-degrade-over-time-6705747/.

[3] Zhang, Y., Tang, Q., Zhang, Y., Wang, J., Stimming, U., & Lee, A. A. (2020). Identifying degradation patterns of lithium ion batteries from impedance spectroscopy using machine learning. Nature Communications, 11(1). https://doi.org/10.1038/s41467-020-15235-7

[4] University of Cambridge. (2020, April 6). AI techniques used to improve battery health and   safety. ScienceDaily. Retrieved April 18, 2020 from    www.sciencedaily.com/releases/2020/04/200406092833.htm

[5] Sit, H. (2019, October 20). Quick Start to Gaussian Process Regression. Medium. https://towardsdatascience.com/quick-start-to-gaussian-process-regression-36d838810319.

Bookmark the permalink.

Leave a Reply

Your email address will not be published.