Joint Estimation of Lithium-Ion Battery Health Status and Remaining Service Life by Transfer Learning Based on PatchTST and Dynamic Weighted MSE Loss Function

Abstract

This study proposes a transfer learning estimation method based on dynamic weighted kernel MSE (DWKMSE) loss function and PatchTST model for the joint estimation of lithium-ion battery health state (SOH) and remaining useful life (RUL). The PatchTST model divides battery aging characteristics into independent features through channel-independent operations, sharing the parameter weights and biases of the transformer backbone to reduce information redundancy and capture key information in each aging feature. The dynamic weighted kernel MSE loss function guides the PatchTST model to update parameter weights, enabling the model to fully learn the nonlinear characteristics of the degradation process and reduce the impact of outliers on the model during training. The effectiveness of the PatchTST model and DWKMSE loss function in the joint estimation of battery SOH and RUL was verified on different battery aging data sets. Finally, transfer learning was performed on two different battery aging data sets to validate the estimation performance of the proposed method under different usage conditions and materials. The experimental index showed that the average MAE value for SOH is 0.421, with an average R² value of 0.953; the average MAE value for RUL is 16.788, with an average R² value of 0.987. Experimental results show that compared with direct training methods, the MAE metric for SOH estimation based on transfer learning decreased by 17.1%, while the R² metric improved by 2.3%; the MAE metric for SOH estimation decreased by 18.6%, and the R² metric improved by 0.1%.