Speaker
Description
After decades of research, four-bar linkages are still an active research topic in engineering, as they can convert a rotary motion into a linear or other, more complex motion if carefully designed. One of the central theorems relating to four-bar linkages is Grashof's law. Based on the four given link lengths, this law dictates the resulting mechanism type and its behaviour, e.g., crank-rocker, double-crank or double-rocker mechanism. In particular, Grashof's law must be fulfilled to obtain a functioning mechanism.
For the rapid development of task-specific mechanisms, the inverse problem must be learned, i.e.: Given a mechanism path, what are appropriate mechanism parameters to produce a similar motion? In recent years, the application of data-driven methods to the design and dimensional synthesis task of four-bar linkages has yielded impressive results. However, when predicting the link lengths of a four-bar linkage, a neural network can also propose infeasible mechanism configurations, especially when the given input is far from the patterns of the training distribution. This raises interest in neural networks that realize prescribed constraints, such that the outputs always guarantee the fulfilment of certain physical conditions, leading to more robust predictions and a better network performance.
We present a novel neural network architecture that incorporates Grashof's law directly into the final layers of the neural network, ensuring that the law is always satisfied for any predicted link lengths. The network thus guarantees the validity of the proposed designs. To evaluate the robustness and network accuracy in the presence of outliers, we compare our proposed method against a naive feedforward MLP approach for paths from a test set that matches the training distribution, as well as with hand-drawn paths for which there may not exist a mechanism that can reproduce the motion exactly. Furthermore, we investigate the impact of different dataset sizes on the performance of the presented architecture.
