A Transfer Learning Approach for Remaining Useful Life Estimation of Lithium-Ion Batteries

Industry 4.0 has reimagined how businesses manufacture and distribute their products. To advance further towards sustainability, the field of Fault Diagnosis and Prognosis (FDP) assumes great significance. The ability to predict failures or to know when a component will reach the end of its operational life can significantly mitigate maintenance and replacement costs. Machine Learning (ML) methods have exhibited the ability to extract trends and models from complex datasets, becoming well-suited for FDP tasks. When dealing with FDP, one of the main problems is the difficulty of obtaining large datasets, due to the burden of conducting extensive laboratory tests, and their usual unbalance, for the impossibility of simulating every possible anomaly that could ever happen. Being able to generate new synthetic data, or to adapt a pre-trained model to another similar task, becomes of paramount importance. In this paper, we focus on lithium-ion batteries. Several commonly used devices are usually powered using lithium-ion batteries, and each of these batteries can assume a completely different behavior from its peers based on usage, charging, and many other factors, leading to potential harm, unreliableness, and other major potential issues. We propose a convolutional Long Short-Term Memory (LSTM) neural network with attention for estimating Remaining Useful Life (RUL) of lithium-ion batteries. The model will be trained on a source dataset, and then re-trained on a smaller target dataset to establish the possibility of applying domain adaptation and transfer learning to RUL estimation, allowing for fast deployment and cost reduction in the production phase. Results show that the use of transfer learning helps to increase the performance of the model, obtaining on the target dataset an accuracy similar to that of the source dataset.