LDNet-RUL: Lightweight Deformable Neural Network for Remaining Useful Life Prognostics of Lithium-Ion Batteries

Deep neural networks (DNNs)-based battery remaining useful life (RUL) prognostics models suffer from high computational costs, long inference times, and a tendency to overfit, posing significant challenges for deployment on memory-constrained edge devices. To address this, LDNet-RUL, a lightweight yet high-performance model, is proposed for lithium-ion battery lifetime prognostics. Specifically, a deformable depthwise convolution unit is designed that efficiently adapts to variations in input features with relatively low computational cost, which leverages fewer parameters and an adaptive receptive field to capture battery degradation dynamics. A multiscale feature extraction method is introduced to extract degradation features from both short-term and long-term data during the battery charging process, which makes the method applicable to various battery charging protocols. To address overfitting and accelerate model inference, an adaptive channel selector, an adaptive residual connector, and a progressive dimensionality-grouped bottleneck filter are incorporated to refine feature selection and reduce the model's complexity, boosting the model's overall performance. Experiments on vehicle lithium-ion battery datasets demonstrate that LDNet-RUL achieves state-of-the-art performance, with average root mean squared error (RMSE) of 0.23 and Score of 0.49. The model is lightweight, with only 7.92 K parameters, achieves inference times of 3.94 ms on GPU and CPU, and 5.39 ms on the NVIDIA Jetson TX2.