Wavelet Transformer-Based Multi-Channel and Multi-Resolution Information Perceptron for Lithium-ion Battery State of Health Estimation

In the application of lithium-ion (Li-ion) battery, indicators like state of health (SOH) are widely adopted in the battery monitoring system. However, it is hard for SOH estimation to be directly obtained from measurement, and it can be interrupted by frequent disturbances and unknown noises. Therefore, in order to accurately predict SOH values, this article proposes a wavelet convolutional encoded deformable transformer (WCED-Trans). WCED-Trans aims to extract frequency battery features in expanded scales and guide the study of model for key information with high relevance, thus increasing the estimation accuracy and alleviating the impact of noises. This model first processes battery signals with multiple discrete wavelet transforms (DWTs) and 1-demensional convolutional neural networks (CNNs), where the frequency features are encoded and transmitted in a multichannel and multiresolution pattern. Then, the encoded data representations are perceived by a deformable self-attention (DSA)-based Transformer encoder. Unlike the traditional Transformer, this encoder introduces a set of offset networks to the self-attention module. The offset groups can refine the study of model by changing the shape of the receptive field with deformed points. The model was experimented on NASA and CALCE dataset, from which the effectiveness and efficiency of this model were proved.