Rapid remaining-useful-life prediction of Li-ion batteries using image-based machine learning

Abstract

With the increased integration of lithium-ion batteries in our everyday lives, accurate and reliable battery management systems have become an imperative aspect of the well-being of our everyday electronics. This thesis proposes the use of novel machine learning methodologies to predict the remaining-useful-life (RUL) of lithium-ion batteries reliably, accurately, and swiftly. Firstly, a method that prides itself on being publicly available, and which can be easily implemented alongside existing methodology, is proposed to increase the prediction accuracy of the conventional health indicator methodology by 6.72% by using images of data curves as inputs. Subsequently, a more in-depth machine learning model is presented which managed to considerably outperform the current literature in terms of speed, accuracy, and reliability, achieving an RUL prediction accuracy of 90.85%. These proposed methodologies have a wide range of applications, from fault diagnostics, state-of-charge, and state-of-health prediction, to other, more complex, regression applications.