Design of a Convolutional Neural Network for the Synthesis of a Spherical Crank-Rocker Mechanism

The use of CNN has proven to be a crucial tool in solving complex problems, particularly since they utilize images as inputs. Given that the synthesis of mechanisms is a complex task, it has been approached in various ways, ranging from solving the equations that describe kinematic motion to methods that employ heuristic algorithms to solve the problem, including numerical methods among others. This work presents an innovative method for the dimensional synthesis of spherical mechanisms, which is based on the use of CNNs. In the proposed method, the trajectory that must be followed by the coupler point of a spherical four-bar mechanism is converted into images so they can be used as inputs for the CNN. The CNN is trained through a database generated with already known spherical mechanisms that met constraints such as the Grashof criterion and transmission angle constraints, among others, since these ensured the mechanism would function properly. In addition, the kinematic study of the spherical mechanism is presented through relative angles, as this allows for the generation of databases and the extraction of kinematic parameters to label the input images used in CNN training. Once the CNN was trained, the neural network was capable of performing the dimensional synthesis of the mechanism. Moreover, it was tested with three different case studies, resulting in three distinct mechanisms. The obtained results were evaluated using the metr2ev function, which allows for the comparison of two trajectories and provides information on their similarity in both time and frequency. The best result obtained was a similarity of 0.9953, indicating that the trajectory generated by the mechanism is reliable concerning the required precision.