An adaptive data augmentation-based reliability evaluation and analysis of lithium-ion batteries considering significant inconsistency in degradation

Lithium-ion batteries (LIBs) exhibit significant degradation inconsistency relative to one another during actual operation. This inconsistency necessitates a large sample size, which is typically impractical from an experimental standpoint, to provide a reliable evaluation of LIB performance. Therefore, this study proposes an adaptive data augmentation-based reliability evaluation and analysis method that dynamically expands the sample size through adaptive data augmentation, ensuring accurate evaluation to the maximum extent. The proposed method integrates a nonlinear Wiener process and Gaussian kernel density estimation to estimate battery lifetime. In addition, an Adaptive Augmentation Magnitudes Generative Adversarial Network (AAM-GAN) algorithm is developed to expand the capacity degradation data until the reliability indices stabilize. AAM-GAN adaptively modifies augmentation magnitudes during training by incorporating Reinforcement Learning, reducing the risks of overfitting and underfitting associated with traditional random-augmentation strategies. The Delta method is applied to quantify the uncertainty of the reliability indices. Experimental results based on 124 LIBs indicated that AAM-GAN reduces the required augmented sample size by 33.3 % and 25.0 % and improves reliability evaluation accuracy by 14.5 % and 7.52 %, compared to the traditional GAN and WGAN-GP, respectively.