Accurate, practical, and robust evaluation of the battery state of health is crucial to the efficient and reliable operation of electric vehicles. However, the limited availability of large-scale, high-quality field data hinders the development of the battery management system for state of health estimation, lifetime prediction, and fault detection in various applications. In this work, to gain insights into underlying factors limiting battery management system performance in real-world vehicles, we analyze the operational data of 300 diverse electric vehicles over three years to understand the disparities between field data and laboratory battery test data and their effect on state of health estimation. Furthermore, we propose a deep learning-based multi-modal framework to effectively leverage historical vehicle data for efficient, accurate, and cost-effective state of health estimation. The proposed paradigm exhibits considerable potential for numerous applications in state estimation and diagnostics in multi-sensor systems. Furthermore, we make the field data of these electric vehicles publicly available aiming to promote further research on the development of effective and reliable battery management systems for real-world vehicles.

The successful integration of statistical machine learning techniques into battery health diagnosis has significantly advanced the development of transportation electrification. To achieve predictive maintenance of batteries, we propose a comprehensive data-driven approach for battery capacity trajectory prediction based on degradation pattern (DP) recognition and health indicators (HIs) extrapolation. First, two HIs ( Qmean/RVmean ) are extracted from 10-minute sequence data before and after a full charge. Second, an unsupervised learning approach is employed for the early-stage battery DP analysis and clustering. Finally, a long short-term memory (LSTM) network is utilized to construct the HIs extrapolation and capacity prediction models. Multi-task learning (MTL) is implemented to predict HI sequences, enabling simultaneous extrapolation of multiple HIs and the sharing of parameters between different HIs during training. A probabilistic neural network is incorporated into the capacity trajectory prediction model to assess the uncertainty of prediction results. The proposed approach is validated using two battery datasets, where predictions for three represented aging stages are presented and evaluated. The results demonstrate accurate and robust predictions, with the average mean absolute percentage error (MAPE) for LiNi _0.86 Co _0.11 Al _0.03 O ₂ (NCA) cells and LiNi _0.83 Co _0.11 Mn _0.07 O ₂ (NCM) cells across all aging stages below 2.59% and 1.15%, respectively.