Edge-based Battery Remaining Useful Life Estimation Using Deep Learning

Lithium-ion batteries are the most commonly used type of battery today that they are used from tiny electronic devices to large electric vehicles. As a result, there has been increasing attention on remaining useful life (RUL) estimation for lithium-ion batteries to predict their failure and ensure their safe operation. However, most research on battery RUL do not test their models to run on low-power edge nodes which are the logical target devices. In this paper, we developed a deep learning-based approach called Edge-Based RUL Estimation (EBRULE). EBRULE performs battery RUL prediction on an edge device. Four deep neural network architectures were trained on two data sets namely the NASA and CALCE datasets, using a GPU, and inference was done on a Raspberry Pi single board computer. The experimental results show that the models were able to make predictions with accuracy of 89% or more within seconds. Our work achieved 0.02-6.81% of RMSE and 0.8-5.67% of MAE. Moreover, the DeTransformer model improves upon the mean absolute error (MAE) and root mean-square error (RMSE) by as much as 71.95% and 77.31% respectively compared to the baseline DeTransformer method while the RNN model shows a reduction in MAE and RMSE by 78.89% and 90.77% respectively compared to the baseline RNN approach.