Comparing Hybrid Approaches of Deep Learning for Remaining Useful Life Prognostic of Lithium-Ion Batteries

Many published journals used hybrid deep learning methods to predict batteries' remaining useful life by adopting different rationales to select and combine deep learning methods aiming to propose the most accurate prediction model possible. The main contribution of this article consists of proposing, to the best of the authors' knowledge, the most accurate hybrid deep learning prediction model, designed and configured by considering the theoretical strength of each of the selected deep learning models, combined with meticulous data preprocessing and feature engineering steps. A benchmark study is presented to confirm the theoretical design by comparing the prediction results of the selected hybrid model with other proposed hybrid deep learning algorithms. The selected prediction model is compared as well with previously published articles, specifically, the ones that have used hybrid deep learning methods, NASA datasets, and batteries #6, #7, and #18 selectively. The hybrid model refers to the combination of different types of deep learning architectures, such as Convolutional Neural Networks (CNNs), Deep Neural Networks (DNN), Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (bLSTM), recurrent neural network (RNN), Bidirectional recurrent neural network (bRNN), Gated recurrent units (GRU) and Bidirectional Gated recurrent units (bGRU). This combination includes CNN-LSTM-DNN, CNN-bLSTM-DNN, CNN-GRU-DNN, CNN-bGRU-DNN, CNN-RNN-DNN, and CNN-bRNN-DNN, and aims to leverage the strengths of each architecture in capturing spatial, temporal, and sequential patterns present in the battery dataset. The hybrid deep learning approaches are tested with multichannel inputs, encompassing parameters such as voltage, current, and temperature, as well as their respective time series averages. The objective is to predict the remaining useful life. Performance evaluation is conducted using error metrics, including Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). The results revealed a remarkable 90.5% enhancement in RMSE, indicating substantial improvement.