A novel RUL prediction framework based on the adaptability feature perception fusion model method

Abstract

To ensure the reliability and safety of lithium-ion batteries operation, accurate prediction of its remaining useful life(RUL) can grasp the internal performance degradation status of the battery in real time and reduce the risk of battery use. A novel RUL prediction framework of variational mode decomposition based on the adaptability feature perception fusion model(AFPFM) is proposed in this paper. Firstly, multiple indirect health indicators (HI) are extracted from the current, voltage, and temperature curves of lithium-ion batteries, and Spearman correlation coefficient method is used to analyze the correlation between HI and capacity. The six health indicators with the highest correlation, namely equal voltage drop discharge time, constant current discharge time, peak discharge temperature, equal voltage difference charging time, equal voltage difference charging energy, and constant current charging time, are selected for RUL prediction. Secondly, a variable mode decomposition(VMD) method is proposed, which decomposes the RUL attenuation curve of lithium-ion batteries into capacity degradation trend component and capacity regeneration component, in order to avoid local fluctuations in capacity regeneration and interference from test noise on RUL prediction results. Thirdly, the RUL prediction framework based on AFPFM is proposed. Capacity degradation is approximately linear, and a linear regression model is proposed to predict the trend of capacity degradation. Combined with the iTransformer model, a transposed Transformer model is proposed to predict capacity regeneration. The adaptability of this model is demonstrated by its ability to be applied to lithium-ion battery capacity degradation datasets of any dimension, the feature perception is achieved by learning attention weights between different HIs through attention mechanisms,and to obtain the final RUL prediction value by fusion the results of two models. Finally, to validate the effectiveness of the method proposed in this paper, the prediction accuracy of proposed model is compared with other commonly time-series prediction models on the NASA and CALCE battery datasets. The results indicate that the proposed model has better RUL prediction performance. For battery B05 dataset, the RMSE, MAE, and MAPE of RUL prediction based on proposed model are 1.94 %, 1.72 %, and 1.27 %, respectively, while the RMSE, MAE, and MAPE of CS2–35 battery are 0.59 %, 0.46 %, and 0.49 %, respectively.