A Fuel Cell’s Big Brother: Artificial Intelligence for Monitoring Fuel Cells

The operation of fuel cell stacks on test benches today is typically monitored by constant alarm threshold values for selected operating conditions. However, due to the wide range of operating conditions of the stacks, this type of monitoring only works for extreme maximum and minimum operating conditions. Wide setting ranges for thresholds prevent the detection of minor faults, whereas too narrow limits will interrupt the tests unnecessarily. A particular challenge is the monitoring of faults that either do not result in a directly measured response from the fuel-cell stack or that only become noticeable with a time delay. The continuous improvement of fuel cell stacks over the last years necessarily requires much-improved monitoring methods, especially for durability tests with an operating time of thousands of hours. For this reason, a novel monitoring concept using AI-based methods for the operation of PEM fuel cells on test benches was developed and is being presented. Firstly, machine-learning based mechanisms for monitoring the operating conditions as set and controlled by the test bench are shown. Due to the cyclic operation during durability tests, the operating conditions set at a load point can be compared to the past operating conditions at the same load point. This proceeding allows the early and accurate detection of faults caused by the test bench and thereby ensures the usability of the measured data as well as an early alarm in case of problems. In addition, a digital twin based method for monitoring the condition of the fuel cell stack is presented. The deep-learning based digital twin calculates a probabilistic prediction of the expected voltage of the fuel cell stack based on the current and past operating conditions. The comparison of expected and measured cell voltage taking the model’s confidence into account enables the early and precise detection of unforeseen events such as contamination and thus averts consequential damage to the fuel cell stack. In contrast to physical models, the digital twin represents a data driven model-ling approach for fuel cells. The presented methods are applied to real testing data to demonstrate the detection of faults during the operation of fuel cell stacks on test benches that would have remained undiscovered by today’s monitoring mechanisms. It shows that the data driven digital twin is able to predict the fuel cell’s stack voltage with an accuracy of 2.5 mV over 1000 h of unseen data. Figure 1