Group equivariant networks for leakage detection in vacuum bagging

The incorporation of prior knowledge into the ma-chine learning pipeline is subject of informed machine learning. Spatial invariances constitute a class of prior knowledge that can be taken into account especially in the design of model architectures or through virtual training examples. In this contribution, we investigate fully connected neural network architectures that are equivariant with respect to the dihedral group of order eight. This is practically motivated by the application of leakage detection in vacuum bagging which plays an important role in the manufacturing of fiber composite components. Our approach for the derivation of an equivariant architecture is constructive and transferable to other symmetry groups. It starts from a standard network architecture and results in a specific kind of weight sharing in each layer. In numerical experiments, we compare equivariant and standard networks on a novel leakage detection dataset. Our results indicate that group equivariant networks can capture the application specific prior knowledge much better than standard networks, even if the latter are trained on augmented data.