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

Abstract

The operation of fuel cell stacks on test benches today is typically monitored by constantalarm threshold values for selected operating conditions. However, due to the wide range ofoperating conditions of the stacks, this type of monitoring only works for extreme maximumand minimum operating conditions. Wide setting ranges for thresholds prevent the detectionof minor faults, whereas too narrow limits will interrupt the tests unnecessarily. A particularchallenge is the monitoring of faults that either do not result in a directly measured responsefrom the fuel-cell stack or that only become noticeable with a time delay. The continuousimprovement of fuel cell stacks over the last years necessarily requires much-improvedmonitoring methods, especially for durability tests with an operating time of thousands ofhours. For this reason, a novel monitoring concept using AI-based methods for the operationof PEM fuel cells on test benches was developed and is being presented.Firstly, machine-learning based mechanisms for monitoring the operating conditions as setand controlled by the test bench are shown. Due to the cyclic operation during durabilitytests, the operating conditions set at a load point can be compared to the past operatingconditions at the same load point. This proceeding allows the early and accurate detection offaults caused by the test bench and thereby ensures the usability of the measured data aswell as an early alarm in case of problems.In addition, a digital twin based method for monitoring the condition of the fuel cell stack ispresented. The deep-learning based digital twin calculates a probabilistic prediction of theexpected 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 intoaccount enables the early and precise detection of unforeseen events such as contaminationand 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 faultsduring the operation of fuel cell stacks on test benches that would have remainedundiscovered by today’s monitoring mechanisms. It shows that the data driven digital twin isable to predict the fuel cell’s stack voltage with an accuracy of 2.5 mV over 1000 h of unseendata. Figure 1