Explainable AI-Based Battery Management System with Federated Learning for Multi-Chemistry State-of-Health Estimation in Heterogeneous EV Fleets

The increased adoption rate of electric vehicles strengthens the demand for efficient battery management solutions with precise state-of-health estimation and support for data privacy within various heterogeneous fleet setups. This work presents a framework for explainable artificial intelligence with the application of federated learning for battery state-of-health prediction with support for privacy preservation within distributed electric vehicle fleets. This work presents a combination model with Long Short-Term Memory networks and the FedAVG learning approach for collaborative model training on ten distributed participants without any central aggregation. With the application of the MIT-Stanford battery dataset with a total number of 124 Li-ion battery cells, the system provides a root mean squared error with a measure of 1.92%, a mean absolute error of 1.49%, and a coefficient of determination with a measure of 0.826 on equal and proper benchmarks while providing complete privacy and local support within all datasets. Additionally, this system comes with a window-based feature engineering approach with dual normalization and complete evaluation for system efficiency on non-identically distributed datasets. The core contributions for this work come with the union and integration between battery health monitoring and distributed federative learning for complete management system support with sub-2% prediction accuracy at reduced prices and from aggregated datasets.