Synthesizing Spatial RSCR Mechanisms for Path Generation Using a Deep Neural Network

Abstract In the recent past, various machine learning approaches have been broadly studied and successfully applied in the planar mechanism synthesis problem. This paves the way for its adoption in the spatial mechanism synthesis that is typically more complex to solve due to dimensionality of the input, large number of mechanism parameters, and highly non-linear relationship between the input and output. This paper introduces a neural network approach for the path synthesis of spatial RSCR mechanisms. We detail the method of database generation with the development of a simulator, data space selection strategies, and a complete data normalization method. A novel way of calculating the coupler curve is utilized, addressing the challenge of describing the absolute position of a point on the rigid body in ℝ3. The database generation uses a spatial mesh grid-base method with specific criteria for preventing circuit defects and duplicate mechanisms. All data in the database is translated, rotated, and reflected to a normalized position. In the end, the classical neural network structure, multi-layer perceptron (MLP), is applied for finding the mapping from the spatial path to mechanism parameters. The MLP with different parameters and techniques, such as batch normalization and dropout, are implemented for finding the best model. The result shows the capability of MLP for complex spatial path synthesis. Moreover, despite the limitation of a single MLP’s structure, this study demonstrates the potential of improvement with more advanced neural networks.