Privacy preserving federated learning for proton exchange membrane fuel cell

Abstract

Proton Exchange Membrane Fuel Cell (PEMFC) is a promising clean energy device with applications from mobile power stations to electric vehicles. To accelerate the application process, deep learning (DL) has been applied to develop various intelligent technologies for PEMFC such as performance prediction, fault diagnosis, etc., to reduce manufacturing cost and prolong service lifetime. However, a single research institution is difficult to obtain sufficient training data for developing DL-based models, since fuel cell system is still in the development stage, and its high cost makes the collection of experimental data too expensive. To tackle the challenges, this study designs a privacy-preserving federated learning framework for PEMFC (FedFC). The framework can support multiple research institutions to collaboratively train a high-performance DL model for PEMFC while preserving their local data information using homomorphic encryption and differential privacy technologies. The study empirically evaluates FedFC framework performance on real fuel cell datasets with performance predication and fault diagnosis tasks. Experiment results demonstrate that the FedFC framework can achieve excellent performance and holds promise for promoting the development of intelligent models associated with PEMFC.