Multi-Task Learning Framework for Fault Detection in Energy Storage System Lithium-Ion Batteries: From Degradation to Slight Overcharge

Abstract

Assessing the state of health (SOH) of batteries and detecting overcharging are crucial in the development of electrochemical energy storage systems. This study proposes a multi-task learning (MTL) framework, specifically devised for application in large-scale energy storage stations (ESS), which is adept at both conducting SOH assessments and identifying slight overcharges. To adapt to real-world operational conditions, we designed aging and slight overcharging experiments to form datasets for the training and evaluation of the MTL framework. The Bi-GRU network was used within this framework, which presented significant advantages in typical scenarios of energy storage stations. Furthermore, modulating the task weights for SOH assessment and slight overcharge detection significantly enhanced the effectiveness of the MTL framework, with a 4:6 weight ratio determined to be ideal for the objectives of this research. The adoption of a hard parameter-sharing architecture markedly diminishes both the inference time and memory demands relative to soft parameter sharing. We implemented a tailored approach for voltage segment input to ensure system safety while enhancing its efficiency. In future works, we aim to detect more types of faults and enlarge the dataset to encompass a broader range of scenarios.

Keywords: Energy Storage Station, Fault Detection, Multi-task Learning, Data Augmentation