Lithium-ion battery application time series data augmentation based on generative adversarial network for training deep learning algorithm

In this paper, a safety accident data augmentation method is proposed to determine the battery failure situation. For data augmentation, overcharging battery electrical abuse accident data is used. A black box-based generative advertising network (GAN) is applied to increase the time series of battery failure data. The effectiveness of the augmented algorithm designed by comparing the time series patterns of overdischarge data generated through GAN with the actual overcharged data is demonstrated. Finally, it is verified that the GAN-based data augmentation method that may be used for abnormality detection of the battery management system (BMS) may be efficiently applied.

I. Introduction

In the realm of modern energy systems, lithium-ion batteries stand out as a cornerstone technology, revolutionizing the landscape of power sources for a multitude of applications. Their widespread adoption can be attributed to a host of advantages, ranging from their eco-friendliness to their remarkable energy density. As versatile reversible conversion devices, batteries play a pivotal role in driving applications through discharge while facilitating the replenishment of energy reserves via charging mechanisms. However, amidst the widespread embrace of lithium-ion batteries, a shadow of concern looms large over the realm of battery-based applications, particularly in the wake of safety incidents such as electric vehicle (EV) fires and energy storage system (ESS) fires, notably observed in regions like Korea. These incidents underscore the critical imperative for proactive measures aimed at enhancing the safety and reliability of battery- powered systems. In response to these challenges, the scientific community has fervently pursued the development of predictive and diagnostic techniques tailored to mitigate the risks associated with battery safety. Central to these efforts lies the utilization of time series data, comprising measurements of crucial parameters such as current, voltage, and temperature, collected over temporal intervals. These data streams offer invaluable insights into the dynamic behavior of batteries, enabling the diagnosis and prediction of transient states within applications. Yet, despite the wealth of research leveraging time series data, a glaring obstacle persists: the scarcity of data pertaining to safety accidents. This deficiency poses a significant challenge, as it introduces an inherent imbalance between normal operational data and data reflecting adverse events. The resultant data skew can lead to detrimental outcomes such as overfitting in predictive models, thereby undermining their effectiveness and reliability. To address this critical gap, this study endeavors to pioneer a novel time series data augmentation technique designed to generate synthetic battery safety accident data. By harnessing the power of generative adversarial networks (GANs), a sophisticated black box-based machine learning framework, the research seeks to overcome the limitations imposed by data scarcity and imbalance.