Predictive Maintenance of fuel cells

As a result of the national and European hydrogen strategy, in which hydrogen is the key to achieve the climate targets, fuel cells will also be used as a powertrain technology in the field of mobility in the future. However, fuel cells are very sensitive systems that can lose considerable performance even if there is only a slight deviation from the optimum operating point. This can result in irreparable damage that greatly reduces the remaining service life. A total loss of the fuel cell even goes hand in hand with a total loss of the vehicle, since the fuel cell is not intended as a replacement component.

Early detection of malfunctions in the form of predictive maintenance for fuel cells increases the lifetime and safety of the fuel cell and offers the possibility to perform individual optimizations. Fraunhofer IMS is in contact with fuel cell manufacturers and system developers to realize innovative solutions for predictive maintenance for fuel cells using embedded AI.

The fuel cell is a complex system and its remaining service life can be significantly reduced by malfunctions and faults. These errors are currently only detected during maintenance or when there is a significant reduction in performance. By then, however, it may be too late and the fuel cell will be permanently damaged and its lifetime reduced. This problem becomes solvable with the help of AI-based systems for predictive failure detection. Furthermore, based on the behavior of the fuel cell, a forecast is possible, which can be used to optimize the operating parameters.

Here, Fraunhofer IMS offers competences in the field of feature extraction  for neural networks and the AIfES library, to realize power-saving, small and fast systems for monitoring the fuel cell. The data is processed directly on the system in or at the fuel cell and does not require an internet or cloud connection: perfect for mobile applications!

Fraunhofer IMS develops neural networks that are as small as possible by extracting crucial features in advance. The minimal size of the neural network increases the traceability of the data processing (greybox approach) and thus supports the use of neural networks also in security-sensitive applications.

However, this principle is not limited to fuel cell applications. An adaptation of predictive maintenance to various other use cases and components in the automobile, such as catalytic converters or turbochargers, is possible. Because the method is platform-independent and already works on small resource-limited systems, it can be used in many places where cost efficiency is very important.