Centralized strategy learning with multi-feature scale health factors for lithium-ion battery RUL prediction

Abstract

The prediction accuracy of the remaining useful life (RUL) of lithium-ion batteries is an important condition to ensure the safe and effective driving of electric vehicles. In the feature factor selection stage, the existing methods mostly use a single health factor, or introduce the method of sample entropy to fuse the features of multiple health factors. This paper proposes a lithium-ion battery RUL prediction based on multi-feature scale health factors and a centralized strategy learning method. It aims to reduce the time complexity of data preprocessing and consider the interaction between data. Firstly, to ensure the richness of the data sequences and its interaction effects, multi-feature scale health factors are extracted from the battery dataset. Secondly, the proposed multi-feature scale star aggregation and redistribution model is adopted. Multiple correlation coefficients were used to analyze the correlation of the extracted data sequences, and then self-adaptive selection was performed. A centralized strategy learning approach was employed to capture inter-sequence dependencies, followed by feature aggregation and redistribution for prediction. Finally, the proposed method was rigorously evaluated on three distinct datasets: the NASA benchmark dataset, the Oxford battery degradation dataset, and our proprietary experimental dataset. Experimental validation demonstrates that the proposed prediction method achieves superior performance, maintaining consistently high coefficients of determination for RUL estimation across all datasets. Compared with other prediction methods, the proposed approach maintains errors below 1.7 %, demonstrating superior accuracy and enhanced generalization capability.