PEM Fuel Cell Fault Diagnosis in Automotive Applications Under Aging Conditions: A Physics-Informed Deep Learning Method

To unlock a massive deployment of proton exchange membrane (PEM) fuel cells in electrified transportation applications it is essential to improve the durability of the technology. Achieving this requires accurate diagnosis of all the faulty operating conditions occurring in the system, which limit the lifespan. In this work, we propose a PEM fuel cell fault diagnosis deep learning (DL) method that integrates physics-based simulations as prior information. This alleviates the need for massive experimental datasets while enhancing the accuracy and the generalization of the algorithm. In the framework of domain adaptation, a supervised adversarial neural network is used to align the features extracted from the simulated data and the experimental data, while making them discriminative to the PEM fuel cell faults. We evaluate the performance of the method using a real experimental dataset collected on a PEM fuel cell operated during 1000 h. Our results demonstrate that integrating physics information in the training data of a fault diagnosis algorithm is highly beneficial for several DL architectures, particularly when limited data are available. A high accuracy level is reached in diagnosing the faults, even with aged data.