Convolutional Transformer-Based Multiview Information Perception Framework for Lithium-Ion Battery State-of-Health Estimation

As a momentous indicator for the working status of lithium-ion batteries, state of health (SOH) has been adopted by various new energy enterprises. Especially in the electric vehicle industry, SOH is an essential guarantee for the safety and robustness of operating batteries. However, working in a complex open environment, the estimation model is often confronted with dynamic changes and abrupt disturbs, which demands an outstanding global information perception ability. Therefore, this study proposes a convolution transformer-based multiview information perception framework (MVIP-Trans) for SOH estimation. MVIP-Trans is dedicated to integrating the benefits of both transformer and convolutional neural network (CNN) architectures to learn global and local features with multiview, thus achieving a comprehensive information perception. First, a local information perceptron (LIP) based on parallel multiscale attention (PMS-A) is constructed in order to extract features from local detailed information in voltage and current signals. With the multiscale attention block, the perceptron has an excellent feature screening ability, thereby enhancing valuable information and restraining useless noise. Subsequently, to get an effective global information perception, a transformer-attention architecture is introduced to encode the global dependencies between the filtered local features. Via multiview learning mechanism, MVIP-Trans adequately modeled the internal correlation between SOH and the physical signals of batteries. Experiments by working through real operation datasets such as NASA PCoE and Oxford prove that this model can perfectly accomplish the SOH estimating task. During the experiments, MVIP-Trans obtained the best performance on NASA dataset, as RMSE is 0.005 and
$R^{2}$
-score reaches 0.989. On oxford dataset, MVIP-Trans got an RMSE of 0.003 and an
$R^{2}$
-score of 0.997. These results prove that our model has a better performance than other existing methods.