Machine Learning and especially deep neural networks are more and more used in industrial applications, because they offer new possibilities to gain knowledge from the available data. The algorithms therefore should work during the lifetime of the industrial system, to which the algorithms are applied for improving the overall performance of the system. However, over the lifetime of an industrial system or application, data sources are likely to change. A sensor may be damaged and delivers faulty values or a new sensor is added to the system. A classical deep neural network will deliver incorrect outputs in case of a faulty sensor or cannot use the additional information coming from a new sensor. In both cases, a retraining with the new input space would be necessary, which requires effort for data collection and training of the network. To address these challenges, this paper proposes an architecture that combines separate approaches, namely the Fast Gradient Sign Method and Transfer Learning. To evaluate the combined architecture, it is applied to a real-world task of failure detection in the Air Pressure System (APS) of an automobile.