Battery Health Prediction Using Commixtured Multi-Feature Selection and Transfer Learning

The data-driven methods have their limitations such as data-insufficiency and data-discrepancy when deployed in real-onboard applications for highly complex and non-linear applications like lithium-ion battery (LIB) health prognosis. Moreover, the accuracy and robustness of the health prediction depends largely on the feature selection for the machine learning (ML) model. To address these shortcomings, this paper proposes a CNN-FiWr-BiLSTM based state of health (SoH) prediction model that uses Convolutional Neural Network (CNN) to automatically extract features from the battery data while a commixture of filter and wrapper methods (FiWr) eliminates both irrelevant and redundant features to obtain more effective feature set. The bi-directional long short term memory (BiLSTM) network learns the ageing features of LIB over time and finally, the SoH regression prediction is made by fully-connected (FC) layer. Furthermore, considering different degradation trajectories exhibited by the dynamic operating conditions, this paper introduces transfer learning (TL) to improve the generalizability and the prediction performance of the model. With experimental validation across the lifetime data of 08 batteries aged cyclically under variable drive profiles, the proposed framework shows overwhelming effectiveness as compared to CNN and BiLSTM alone with root mean square error (RMSE) of less than 0.14% and maximum absolute error (MAE) of less than 02.99 for all batteries.