Study on Improving Reliability of Synthetic Battery Data with TimeGAN Pre-processing

While eco-friendly energy is attracting the public’s interest in recent environmental issues, the market of lithium-ion batteries is expanding. Ensuring the safety and reliability of batteries is important in using batteries. Hence, it is essential to monitor the state of health (SOH) of the battery. Model-based SOH estimation has a limitation because battery data exhibit non-linear characteristics according to complex internal changes. To overcome this limitation, the data-based estimation of the SOH of the battery has been researched extensively. Data-based SOH estimation requires a large, high-quality battery dataset. In general, a battery dataset is built based on data obtained from experimental tests. However, this approach is time-dependent. To overcome this problem, data augmentation algorithms are used. TimeGAN is one of the time-series data augmentation algorithms that generate time-series data in the latent space while preserving temporal dynamics and static characteristics. To ensure stable training of data, TimeGAN generates synthetic data after data has been preprocessed using Min-Max scaling. In this study, various preprocessing techniques were applied for TimeGAN to generate more realistic synthetic data which was close to the original battery data. Results based on different preprocessing techniques were analyzed, and the preprocessing technique suitable for generating a battery dataset was proposed. Discriminative and predictive scores were used for evaluation with t-SNE graphs for visual comparison. When Max-Abs scaling and Normalization preprocessing techniques were applied, the discriminative score improved by 53.7 and 44.4%, respectively, compared to the conventional Min-Max scaling.