Addressing Data Scarcity in PEMFC Fault Diagnosis Using Adversarial Learning

To pave the way to a large industrial deployment of proton exchange membrane fuel cells (PEMFCs) it is essential to improve the durability of the technology. Achieving this requires accurate diagnosis of abnormal operating conditions. To automate the monitoring process in PEMFC, data-driven fault diagnosis models have shown great potential. However, in practice, the performance of data-driven models can be compromised by the availability of the data for training these algorithms. This article proposes a simulation-driven domain adaptation method to circumvent the data scarcity issue using a physics-based model. Through adversarial training, we focus on extracting relevant, high level features for the fault diagnosis task from the simulated dataset, while simultaneously aligning the features extracted from a real dataset acquired on a PEMFC stack. Introducing supervision on the real dataset, we are able to extract features that are discriminative for the fault diagnosis task as well as invariant to the domain. The experimental results validate the effectiveness of the proposed method in dealing with scenarios of insufficient data availability for PEMFC fault diagnosis.