Monitoring Deep Learning Models in Aviation

What is it about?

Environment perception will play an important role for autonomous aircraft, e.g., to be able to prevent mid-air collisions or to find emergency landing spots. Deep Learning (DL) based approaches for computer vision often give state-of-the-art results but are currently not certifiable for aviation because of their data driven training process and their black-box character. Runtime monitoring of the model input could mitigate this problem by ensuring that the model output is only considered when the input is deemed to be suitable. On the one hand, this could be achieved by monitoring operational parameters described by an Operational Design Domain (ODD) as suggested by the European Union Aviation Safety Agency (EASA). On the other hand, unsafe input data might be rejected based on its direct impact on the model performance using Out-of-Model-Scope (OMS) detection. However, performing either ODD monitoring or OMS detection for high-dimensional input data such as camera images is a non-trivial task as it is unclear which properties of an input image should be monitored. In this work, we describe a process to derive a set of suitable low-level image properties that can be used to monitor the input of a DL component. We show that the features selected by the process can be used by a runtime monitor to improve the safety of a DL component by filtering images that violate the ODD boundaries or are OMS

Featured Image

Photo by Alessio Soggetti on Unsplash

Photo by Alessio Soggetti on Unsplash

Why is it important?

In computer vision deep learning methods often achieve the best results. However, it is hard to certify them for aviation due to their data driven training process and their black-box character. Monitoring the model input during runtime could mitigate this problem by ensuring that the model output is only considered when the input is deemed to be suitable.