Wavelet attention-powered neural network framework with hierarchical dynamic frequency learning for lithium-ion battery state of health prediction

Abstract

Accurately predicting the state of health (SOH) of lithium batteries is vital for tasks such as electronic device monitoring. However, current physical information such as voltage and current measured using sensors has important frequency components, which are difficult to learn by existing models, and frequency aliasing often occurs. Therefore, a wavelet attention-powered hierarchical dynamic frequency learning framework (WAPHF) is proposed. Specifically, a wavelet-powered frequency learning layer (WPFL-Layer) is designed to extend the feature learning space of convolution neural network (CNN) to the frequency domain, which gives CNN the ability to learn frequencies and separate the frequency components of different intervals. In addition, the proposed dynamic frequency-focused attention (DFFA) module gives the framework the ability to focus on valuable frequency features and select them dynamically. WAPHF innovatively integrates CNN and wavelet transform together, fully exploiting the ability of CNN to extract features hierarchically and wavelet transform to analyze frequencies. In a constantly hierarchical and dynamic process of frequency learning, the problem of frequency aliasing can be solved. Experiments on three datasets show that the framework can accurately predict the SOH of batteries and outperforms state-of-the-art methods. In addition, a visual analysis demonstrates how DFFA selects and learns useful frequency features.