A Comparative Analysis Of Deep Learning For Remaining Useful Life Estimation
Main Article Content
Abstract
Predicting Remaining Useful Life (RUL) accurately in when it comes to Lithium-ion (Li-ion) batteries is very important as the demand for it is ever increasing. The maintenance and accurate prediction of the RUL of the batteries is growing as the demand for it is increasing in an exponential order mainly due to the sudden growth in the EV industry. This paper performs a comparative analysis on four of the most widely implemented deep learning architectures for RUL prediction using a dataset developed by the Centre for Advanced Life Cycle Engineering (CALCE). Models that are implemented include Long Short-Term Memory (LSTM) networks, Convolution Neural Networks (CNNs), a hybrid Autoencoder + LSTM architecture, and Transformer based model. These models are evaluated for performance on basis of computed parameters such as MSE and MAE i.e. the Mean Squared Error and Mean Absolute Error. They are also evaluated on the R2 score. This comparison of performance between the models aims to identify the model that predicts the RUL most accurately, contributing to the better battery health management system. The experimental results show us that hybrid models such as the Transformer based model and the LSTM + Autoencoder model have a superior prediction performance in comparison to traditional deep learning models such as LSTMs and CNNs.