Friction stir welding is an advanced joining technology that is particularly suitable for aluminum alloys. Various studies have shown a significant dependence of the welding quality on the welding speed and the rotational speed of the tool. Frequently, an inappropriate setting of these parameters can be detected through an examination of the resulting surface defects, such as increased flash formation or surface galling. In this work, two different learning-based algorithms were applied to improve the surface topography of friction stir welds. For this purpose, the surface topographies of 262 welds, which were performed as part of ten studies, were evaluated offline. The aim was to use reinforcement learning and Bayesian optimization approaches to determine the most appropriate settings for the welding speed and the rotational speed of the tool. The optimization problem was solved using reinforcement learning, specifically value iteration. However, the value iteration algorithm was not efficient, since all actions and states had to be iterated over, i.e., each possible parameter combination had to be evaluated, to find the best policy. Instead, it was better to solve the optimization problem directly using the Bayesian optimization. Two approaches were applied: both an approach in which the information from the other studies was not used and an approach in which the information from the other studies was used. On average, both the Bayesian optimization approaches found suitable welding parameters significantly faster than a random search algorithm, and the latter approach improved the result even further compared with the former approach. Future research will aim to show that optimization of the surface topography also leads to an increase in the ultimate tensile strength.

No abstract

Spiking neural networks offer the potential to drastically reduce energy consumption in edge devices. Unfortunately they are overshadowed by today's common analog neural networks, whose superior backpropagation-based learning algorithms frequently demonstrate superhuman performance on different tasks. The best accuracies in spiking networks are achieved by training analog networks and converting them. Still, during runtime many simulation time steps are needed until they converge. To improve the simulation time we evaluate two inference optimization algorithms and propose an additional method for error minimization. We assess them on Residual Networks of different sizes, up to ResNet101. The combination of all three is evaluated on a large scale with a RetinaNet on the COCO dataset. Our experiments show that all optimization algorithms combined can speed up the inference process by a factor of ten. Additionally, the accuracy loss between the original and the converted network is less than half a percent, which is the lowest on a complex dataset reported to date.