Large-scale field data-based battery aging prediction driven by statistical features and machine learning

Accurately predicting battery aging is critical for mitigating performance degradation during battery usage. While the automotive industry recognizes the importance of utilizing field data for battery performance evaluation and optimization, its practical implementation faces challenges in data collection and the lack of field data-based prognosis methods. To address this, we collect field data from 60 electric vehicles operated for over 4 years and develop a robust data-driven approach for lithium-ion battery aging prediction based on statistical features. The proposed pre-processing methods integrate data cleaning, transformation, and reconstruction. In addition, we introduce multi-level screening techniques to extract statistical features from historical usage behavior. Utilizing machine learning, we accurately predict aging trajectories and worst-lifetime batteries while quantifying prediction uncertainty. This research emphasizes a field data-based framework for battery health management, which not only provides a vital basis for onboard health monitoring and prognosis but also paves the way for battery second-life evaluation scenarios.