Towards Efficient SOH Estimation for Lithium-ion Batteries via Structural Re-parameterization

Online state of health (SOH) estimation is critical for the safety and performance of lithium-ion batteries, especially in dynamic environments like electric vehicles and grid energy storage systems, where accurate monitoring prevents failures and enhances reliability. Convolutional Neural Networks have shown great potentials in SOH estimation. To improve the estimation accuracy, traditional methods focus on increasing the depth and width of the network. The resulting network structure is often complex and difficult to deploy on resource limited mobile devices. Based on the structure reparameterization technique, this paper simplifies the network structure from the training stage to the inference stage. Tuning the hyperparameters of a complex network is not an easy task. In this paper, an advanced hyperparameter optimization engine is used to automatically search for the optimal hyperparameters of the network structure. Finally, in the inference stage, the FLOPs of the model after structure reparameterization are reduced by 16.799% as least. The SOH estimation results on two battery aging datasets demonstrate the superiority and robustness of the proposed model, with a maximum average performance improvement of 21.92%.