View planning in the visual inspection for remanufacturing using supervised- and reinforcement learning approaches

Visual inspection in remanufacturing, despite technological progress, is still mainly performed by humans. A rough assessment of the product’s general condition and the dedicated inspection of individual product features or defects is necessary to identify the typically unknown product variant and assess the reusability of a used product and its components. Therefore, a system for automated visual inspection must be flexible and runtime-adaptive, as defects to be inspected in detail may occur anywhere on the product. In the present work, this problem is framed as a view planning problem solved by means of supervised learning and reinforcement learning using a specially developed simulation environment. Three variants of neural networks (PointNet, PointNet++, and Point Completion Network) are compared in the supervised learning case, whereas a deep learning SAC algorithm using the Point Completion Network as network structure is evaluated in the reinforcement learning case. Considering the specific boundary conditions prevailing in remanufacturing, the results are obtained from the use case of electric starter motor remanufacturing. The results show that supervised learning and reinforcement learning are suitable for determining the poses of an acquisition system at system runtime to react to an initially unknown inspection task.