A framework for Li-ion battery prognosis based on hybrid Bayesian physics-informed neural networks

Area of Science:

• Materials Science
• Electrical Engineering
• Computer Science

Background:

• Lithium-ion batteries (Li-ion) are crucial for electric propulsion, demanding accurate state of charge and health monitoring for reliability.
• Traditional physics-based models are computationally intensive and unsuitable for real-time prognosis and health management.
• Existing machine learning methods often struggle to capture complex electrochemical dynamics and uncertainties.

Purpose of the Study:

• To propose a hybrid physics-informed machine learning (PIML) approach for accurate Li-ion battery prognosis.
• To address the limitations of purely data-driven or physics-based models in battery management systems.
• To enhance the reliability and safety of electric propulsion systems through advanced battery analytics.

Main Methods:

• Implemented a PIML framework using recurrent neural networks (RNNs) for direct numerical integration of governing equations.
• Utilized multi-layer perceptrons (MLPs) to model form uncertainty and variational MLPs for battery-to-battery aleatory uncertainty.
• Employed a Bayesian approach to integrate fleet-wide data (priors) with battery-specific discharge cycles.

Main Results:

• The proposed hybrid PIML model effectively simulates dynamical responses of Li-ion batteries.
• The framework accurately captures both model-form and aleatory uncertainties in battery performance.
• Demonstrated effectiveness using the NASA Prognostics Data Repository Battery dataset.

Conclusions:

• The hybrid PIML approach offers a computationally efficient and accurate solution for Li-ion battery prognosis.
• This method significantly improves the reliability of battery monitoring for electric propulsion applications.
• The framework provides a robust tool for managing battery health across diverse operational conditions.