Investigation and application of machine learning methods for automated fuel cell monitoring

Addressing and mitigating global warming is one of the greatest modern challenges for humanity. Achieving this goal requires a shift from fossil fuels to renewable energy sources. Hydrogen plays a key role in this transition due to its high gravimetric energy density, alongside direct electrification. Electrochemical systems such as electrolyzers and fuel cells enable the conversion of electrical energy into chemical energy (in the form of hydrogen) and vice versa. In particular, hydrogen-powered proton exchange membrane fuel cells (PEMFCs) are a crucial technology for transitioning to renewable energy in heavy-duty mobile applications such as trucks, ships, and airplanes. While PEMFC technology has progressed beyond the development stage and is now entering industrialization, meeting the demanding requirements for durability and reliability, often exceeding 25000 h, remains a significant challenge. Continuous monitoring during operation can help identifying potential faults early, thereby preventing permanent damage and improving both durability and reliability. However, current state-of-the-art threshold-based monitoring methods primarily detect only the most extreme fault patterns, especially during testing. This work, therefore, focuses on developing new monitoring concepts based on machine learning. These methods are validated using test station data, as no real world data was available at the time of the study. Given the complexity of fuel cell test stations and systems, as well as the multitude of possible faults, the monitoring process is divided into two stages. The first stage monitors peripheral components to ensure that operating conditions are suitable for the fuel cell. The second stage focuses on fuel cell monitoring, where faults are identified based on their effects on the cell voltage. Peripheral monitoring detects faults in components such as temperature, pressure, humidity, and flow sensors by identifying deviations in their measurements. A distinction is made between fast-changing parameters (e.g., pressure, current, and mass flow) and slow-changing parameters (e.g., temperature). For fast-changing parameters, faults can often be detected directly via control deviations. For slow-changing parameters, the cyclic operating patterns typical of test station operation is exploited to identify faults specific to each load point. In both cases, simple algorithms, such as a 3-sigma rule, are shown to be sufficient for accurately detecting faults. Additionally, basic physical models enable the detection of discrepancies between related sensor readings, such as pressure drop and flow. A wide range of fault sources, such as contamination of operating media by pollutants, are not directly measured but manifest only through their impact on the fuel cell's voltage. For this reason, cell voltage is the most critical parameter for assessing the fuel cell's state of health. However, since cell voltage is strongly influenced by both operating conditions and temporal effects, all known influences must be accounted for to identify any unknown effects, such as contamination. To achieve this, a digital twin of the fuel cell is developed in this work to predict cell voltage based on current and historical operating data. This prediction is then compared with real measurements to identify anomalies. For the first generation of the digital twin, two machine learning algorithms - feed-forward neural networks (FF-NNs) and long short-term memory (LSTM) neural networks - are evaluated for their predictive accuracy. To improve performance, artificial features, such as the time since the last shutdown, are included in the predictions. The models trained with test station data are then validated through transfer learning by adapting them to new fuel cell stacks. The LSTM model, in particular, demonstrates high predictive accuracy, with deviations of only a few millivolts, enabling early fault detection. However, adapting the model to new fuel cell stacks shows to be computationally intensive and prone to errors. To address these challenges, a second-generation digital twin is developed, consisting of two complementary models. The first model, a FF-NN, predicts the steady-state cell voltage based on current operating conditions. Since temporal effects are excluded, a stationary physical fuel cell model is used to generate synthetic training data. This model can then be fine-tuned to a real fuel cell via transfer learning, requiring only 25 data points to outperform the physical model. In addition, a degradation model based on LSTM is developed to account for all temporal effects. Since this model only computes a correction term for the steady-state model, the influence of most operating conditions can be disregarded, significantly reducing the required amount of training data. Together, the two models achieve predictive accuracy within a few millivolts, with significantly improved stability when transferring the trained models to new fuel cells compared to the first generation. Finally, two systems for automated fuel cell test monitoring are introduced. The first system enables the real-time transfer of test station measurements to a central database. A message broker is used to decouple data acquisition from data archiving while ensuring data integrity. The second system, connected to the database, performs data validation using the developed monitoring concepts. The algorithms are applied to new measurement data and periodically retrained with the validated data. If anomalies are detected, notifications can be sent to relevant personnel.