Early Prediction of Lithium-Ion Battery Lifetime with Random Charge-Discharge Behaviors and Feature Noise

The accuracy of lithium-ion battery lifetime prediction heavily relies on high-quality data, which typically exhibits complete charge-discharge cycles and low feature noise. However, accurate lifetime prediction faces significant challenges when batteries are subject to random charge-discharge behaviors and intrinsic feature noise. To address this issue, this paper proposes an early prediction framework suitable for such conditions. The proposed approach incorporates an adaptive variable window moving average (AVWMA) filter, which utilizes change-point analysis (CPA), along with a dynamic window adjustment strategy to effectively balance curve smoothness and feature preservation. The proposed feature segment extraction method employs a dynamic extraction strategy based on sliding windows and time series similarity measure, enabling effective extraction of degradation-related features from incomplete incremental capacity (IC) curves. A two-dimensional hierarchical feature matrix is then constructed, and a convolutional neural network (CNN) is designed to extract latent degradation information from both spatial and temporal domains for lifetime prediction. Experimental results demonstrate that the proposed framework can effectively mitigate data incompleteness and feature noise, achieving precise early lifetime prediction and enhancing the applicability of such methods in real-world scenarios.