Sequential application of denoising autoencoder and long-short recurrent convolutional network for noise-robust remaining-useful-life prediction framework of lithium-ion batteries

Abstract

Accurate prediction of the lithium-ion battery lifetime is important to maintain the performance of the battery system. Because data-driven methods are extensively used in the analysis of nonlinear dynamical systems in research, the existing literature is largely focused on the application of these methods for the prediction of battery state. Data-driven methods that are highly dependent on data are sensitive to the noise of the measurement data. Distorted data diminish the performance of data-driven models for lifetime prediction. In this study, we propose an artificial neural network-based framework, which is robust to noise, to increase the prediction accuracy of the remaining-useful-life (RUL) of lithium-ion batteries. A denoising autoencoder trained using a distorted dataset with Gaussian noise and dropout is presented to improve the robustness of the model to noise. Artificial neural network models predict RUL based on the state-of-health estimated using measurement data in the initial step. The proposed prediction model is compared with the base model and the training-noise model using distorted data to validate the robustness to noise. The results demonstrate that the proposed framework is robust to noise and has lower cycle errors compared to other methods.