A dual-network approach for avoiding feature ambiguity in the synthesis of crank-driven four-bar linkages

We investigate the synthesis of crank-driven four-bar linkages by learning the inverse problem of predicting the mechanism parameters from given coupler-point paths. This problem exhibits an ambiguity in feature space, where different mechanisms may produce very similar paths. This paper proposes a dual-network approach to combat this issue and compares it to a naive single-network approach. Furthermore, different feature extraction methods, normalizations, and feature lengths are evaluated. We show that the dual-network approach generally leads to better and more robust predictions. Lastly, we show that a neuro-initialized post-optimization can locally refine the proposed designs and achieves results comparable to intermediate solutions of a global optimization algorithm but requires a much lower computation effort.