Domain Adversarial Neural Networks for Remaining Useful Life Estimation of Lithium-Ion Batteries

During the last few years, the number of devices powered by lithium-ion batteries has grown exponentially, and so has the number of accidents caused by this kind of batteries. Lithium-ion batteries can assume a completely different behavior from their peers based on usage, charging, and many other factors, leading to potential harm and other major issues depending on the importance and purpose of the specific device being powered. It stems from this the importance of being able to accurately predict the Remaining Useful Life (RUL) of such batteries, and most importantly creating a model that is capable of generalizing across different sets of batteries. This work introduces a Domain Adversarial Neural Network (DANN) architecture which, using the adversarial learning paradigm, aims to transfer the knowledge on a source dataset to a target dataset, reducing the shift between their features distribution. The DANN offers a distinct advantage over traditional transfer learning methods, as it does not rely on explicit labels from source or target domains, making it particularly valuable in scenarios with limited availability of labeled data. Additionally, its flexibility allows seamless integration into various neural architectures, such as the convolutional-LSTM neural network that we used in our work. For empirical evaluation we used the MIT-Toyota 2019 collaboration dataset, which is the largest lithium-ion battery dataset publicly available, showing the goodness of our method when compared to traditional transfer learning methods.