This paper reports our initial development of a simulation-informed machine learning algorithm for failure diagnostics in solid oxide fuel cell (SOFC) systems. We used physics-based models to simulate electrochemical impedance spectroscopy (EIS) response of a short SOFC stack under normal conditions and under three different failure modes: fuel maldistribution, delamination, oxidant gas crossover to the anode channel. These data were used to train a support vector machine (SVM) model, which is able to detect and differentiate these failures in simulated data under various conditions. The SVM model can also distinguish these failures from simulated uniform degradation that often occurs with long-term operation. These encouraging results are guiding our ongoing efforts to apply EIS as a failure diagnostic for real SOFC cells and short stacks.

In this work, we apply a machine learning approach to solid oxide fuel cell (SOFC) system diagnostics. Instead of fitting electrochemical impedance spectroscopy (EIS) into a physics based model or equivalent circuit, we train machine learning models to recognize failures from a database of simulated EIS. We use a coarse-grained physics-based model to simulate stack EIS under three different failure modes: fuel maldistribution, delamination, and cathode gas crossover to anode channel. Synthesized machine learning classification models successfully recognize these different degradation mechanisms in simulated data across different operating conditions. We are also able to differentiate these failures from the uniform degradation that tends to occur with SOFC over time. These encouraging results prompt our current effort to implement machine learning diagnostics methods on experimental EIS collected on SOFC short stack.

Solid oxide fuel cells (SOFC) offer great potential for efficient use of renewably-generated storage fuels. However, commercialization of SOFCs is currently limited by a lack of long-term durability. This work explores the potential of using electrochemical impedance spectroscopy (EIS) as in-situ diagnostic of SOFC stack and system health based on machine learning. A multi-physics 2D/3D-model of SOFC, operating with humidified hydrogen or natural gas fuel, is used to simulate impedance data at various nominal operating conditions. The model is also used to simulate impedance responses under different critical failure modes, such as uneven fuel distribution that leads to high local fuel utilization, electrode delamination or deactivation. Based on impedance simulations, classification models are synthesized to recognize patterns and distinguish different degradation mechanisms. Based on these models, we believe that impedance measurements can provide better indications of SOFCs system health compared to polarization curves.

Electrochemical energy conversion based on electrolysis and fuel cells offers one possible technological route to long term storage of renewable electricity. However, a major factor limiting commercial deployment is uncertainty about the long-term durability of these technologies in real-world operation. To be successful, these technologies must incorporate ways to pre-detect and mitigate failure. By coupling well-known electrochemical techniques with novel data analysis tools, we believe it is possible to identify key defects and pre-failure conditions in operating fuel cell systems, allowing these systems to be serviced or replaced before catastrophic failure and downstream consequences can occur. Through experimental study on a solid oxide fuel cell (SOFC) short stack of 6 cells, we identify the mechanisms contribute to this system electrochemical impedance spectroscopy (EIS), by carrying out pristine cell/stack tests at different operating conditions (gas partial pressure, current, temperature). Information gathered from the experiments informs a coarse-grained physics-based model, which provides good agreement with experimental data. We use our coarse-grained model to simulate different possible failures in the cell to explain the EIS we collected of failed/degraded cells. We applied a machine learning approach to this process of failure identification by training a model to detect failure on a database of simulated impedance of healthy and failed SOFC. The model robustness is demonstrated as it well predicts experimental data. Figure 1

In this study, we propose a new experimental setup and methodology that can perform in-situ diagnostic of a solid oxide cell stack without interfering with its normal steady state operation. We establish a floating ground galvanodynamic electrochemical impedance spectroscopy measurement setup connected to a 6-cell short stack. The diagnostic setup is electrically connected in parallel to the electronic load of the stack, so that the respective AC and DC current signals overlap. We demonstrate that this methodology is effective in detecting degradation and assessing the state-of-health of a specific cell or stack without disrupting its operation.