Battery Time-Series Data Augmentation Using a GAF-Based Generative Framework

Generative models have gained significant attention in computer vision and natural language processing for their ability to generate realistic samples from complex data distributions. Inspired by these advancements and the persistent challenge of data scarcity in time-series domains, this work proposes a pioneering framework that transforms time-series data (e.g., voltage, current, temperature of lithium-ion batteries) into Gramian Angular Field (GAF) images for synthetic data generation using deep convolutional generative adversarial network (DCGAN) architecture. The proposed GAF-based GAN framework, validated on the Massachusetts Institute of Technology (MIT) battery dataset, generates samples that closely resemble real data and enables accurate signal reconstruction through inverse GAF transformation, effectively overcoming scale invariance and invertibility issues. To further demonstrate the quality of the synthetic signals, state of charge estimation (SOC) is performed using both original and augmented datasets. The results present that using augmented datasets help the models perform better than when trained only on the original data, highlighting the effectiveness of the proposed method for improving time-series data augmentation and state estimation.