Convolutional Attention Transformers for Predicting the Remaining Useful Life of Lithium-Ion Batteries

Reliable estimation of the remaining useful life (RUL) of lithium-ion batteries plays a crucial role in alleviating user concerns about their operational safety and dependability. Despite significant advancements, many current studies concentrate on the degradation patterns of individual batteries, often overlooking the informative value embedded in multiple degradation indicators and their interrelations. To address this gap, we introduce a novel architecture that combines local convolutional attention with a grouped range convolutional attention strategy. In particular, the temporal encoding layer leverages local convolution to extract short-range temporal dependencies, thereby improving the representation of localized temporal dynamics. In the feature encoding layer, grouped convolution is utilized to explore latent interactions among multi-sensor signals, improving the extraction of feature correlations. These two types of representations are subsequently fused in the feature fusion layer to generate more informative and discriminative feature vectors. Moreover, to further enhance the robustness of RUL predictions, an uncertainty-weighted loss function is introduced, allowing the model to dynamically adjust to prediction deviations based on uncertainty estimation. Experimental results on two established lithium-ion battery benchmarks verify that our method delivers marked improvements over current leading techniques in accurately and reliably estimating RUL.