Nested-Mixer: A Nested Deep Learning Model for Online State of Health Prediction of Lithium-Ion Batteries with Multivariable Input

In the task of state of health (SOH) estimation, fully utilizing the multivariate time-series data collected by sensors can bring more accurate results to the battery management system (BMS). However, current work seldom discusses how to integrate information between multiple variables effectively. In this paper, a lithium-ion battery SOH prediction model named Nested-Mixer is designed, which combines convolutional neural networks (CNNs) and Transformers. The time-series data of different variables are divided into different channels as model inputs. CNN Encoder is used to extract features independently from each channel of the input; Intra-Channel Mixer employs a Transformer module with bottleneck attention to effectively extract temporal information within each channel. Inter-Channel Mixer, which is nested with the Intra-Channel Mixer, is used to model the interactions between channels. The proposed model has been tested on two datasets. Results show that Nested-Mixer can automatically extract crosschannel features that are highly correlated with the battery's SOH, achieving higher prediction accuracy and stronger robustness under the same input conditions.