Early‐stage lifetime classification for batteries: Machine‐learned features versus human‐selected features

This study integrates shallow machine learning tools and deep learning algorithms to classify battery lifetime early in the cycling process. We assessed these approaches using human‐selected and data‐driven features across three datasets (A, B, and C). Traditional machine learning models utilized human‐selected features and achieved remarkable classification accuracy. With data from the first two cycles, batteries maintaining over 80% capacity after 550 cycles were classified with 96% accuracy, increasing to 98% as more cycles were included. Even with data from a single cycle, these models achieved 94% accuracy, demonstrating the robustness of human‐selected features. Key contributors included internal resistance, particularly its maximum, average, and minimum values, and features from charging policies such as CC‐CV (constant current‐constant voltage) parameters. Features associated with CC mode were especially significant. Temperature features during charge and discharge also played a role, while capacity fade contributed minimally. The deep learning framework, by contrast, extracted just four data‐driven features to achieve 98.5% accuracy with data from 100 cycles and 99.5% with 200 cycles. Notably, our study explored how classification accuracy trends with varying amounts of data, highlighting critical differences between shallow machine learning and deep learning methods in this respect. While deep learning exhibited impressive accuracies at higher cycles, its performance declined significantly when fewer early cycles were available. Data‐driven features also showed limitations in generalizing across distinct battery profiles, unlike human‐selected features, which demonstrated strong adaptability and broader generalization across different chemistries.