Early lifetime prediction of lithium-ion batteries based on classical image encoding methods

Accurate early-stage lifetime prediction of lithium-ion batteries is critical for ensuring operational safety and improving the efficiency of quality grading. However, it still remains challenging due to the long duration of degradation tests, the limited number of available training samples, and the scarcity of aging features in early-cycle data. Despite Convolution neural networks have shown promise in lifetime prediction, most existing research primally focus on model design and optimization, while neglecting further effective process of input data. In this work, five classical image encoding methods are investigated to transform time-series capacity-voltage sequences into two-dimensional images, with the aim of enhancing subtle degradation features embedded in early-cycle data. A residual neural network is then designed to extract spatial features and predict lifetime from these images. Experiments are conducted to evaluate the performance of different encoding methods under varying input combinations and early-cycle data selections. The results reveal that appropriate encoding methods emphasizing intra-cycle relationships significantly improve prediction accuracy, reducing the root-mean-square error (RMSE) by an average of 15 cycles compared to methods using raw data directly. In particular, the relative position matrix method consistently outperforms the others by uncovering distance-based features within capacity-voltage sequences, achieving an RMSE of 90.2 cycles using only the 10th and 100th cycles data—comparable to models using all 100 cycles data. Additionally, organizing the initial and end states of early-cycle data into dual-channel images achieves better performance than other input combinations.