Abstract
The precise estimation of the Remaining Useful Life (RUL) of lithium-ion batteries is essential for averting unforeseen failures and enhancing operational efficiency and maintenance planning. This paper presents an advanced deep learning framework that couples a spatial-attention mechanism with a Transductive Long Short-Term Memory (TLSTM) model, augmented by one-dimensional dilated convolutional layers to capture long-range temporal dependencies. In contrast to traditional LSTM or GRU models, our methodology utilizes one-dimensional dilated convolutional layers to effectively capture long-range temporal relationships and implements a clustering-based Differential Evolution (DE) strategy for resilient weight initialization and optimization. This hybrid architecture boosts prediction robustness and solves the persistent issue of capacity regeneration effects, frequently neglected by previous models. Additionally, we offer a many-to-one multi-channel learning framework that incorporates several operational signals (voltage, current, temperature) instead of depending exclusively on single-channel capacity data. This invention minimizes parameter complexity while markedly enhancing generalization to unfamiliar settings. Extensive experiments performed on the publicly accessible NASA lithium-ion battery dataset reveal that the proposed model significantly outperforms existing state-of-the-art methods, achieving error reductions of over 10-14% and an exceptionally low Mean Absolute Percentage Error ranging from 0.0053 to 0.0095. These findings distinctly underscore the originality of our research: It represents the inaugural framework that amalgamates spatial attention-driven TLSTM with dilated convolution and clustering-based DE optimization in a multi-channel context, setting a new standard for precise and dependable RUL prediction in battery health monitoring and predictive maintenance applications.