Open Access

Philipp Mascha

• After ground-breaking achievements through the application of modern deep learning, there is a considerable push towards using machine learning systems for visual inspection tasks part of most industrial manufacturing processes. But whilst there exist a lot of successful proof-of-concept implementations, productive use proves problematic. Whilst missing interpretability is one concern, the constant presence of data drift is another. Changes in pre-materials or process and degradation of sensors or product redesigns impose constant change towards statically trained machine learning models. To handle these kind of changes, a measurement of system confidence is needed. Since pure model output probabilities often lack in this concern better solutions are required. In this work, we compare and contrast several pre-existing methods used to describe model confidence. In contrast to previous works, they are evaluated on a large set of real-world manufacturing data. It is shown that utilizingAfter ground-breaking achievements through the application of modern deep learning, there is a considerable push towards using machine learning systems for visual inspection tasks part of most industrial manufacturing processes. But whilst there exist a lot of successful proof-of-concept implementations, productive use proves problematic. Whilst missing interpretability is one concern, the constant presence of data drift is another. Changes in pre-materials or process and degradation of sensors or product redesigns impose constant change towards statically trained machine learning models. To handle these kind of changes, a measurement of system confidence is needed. Since pure model output probabilities often lack in this concern better solutions are required. In this work, we compare and contrast several pre-existing methods used to describe model confidence. In contrast to previous works, they are evaluated on a large set of real-world manufacturing data. It is shown that utilizing an approach based on auto-encoder reconstruction error proves to be most promising in all scenarios tested.…