Selective Assembly of Rolling Element Bearings Based on Pointer Network

Radial clearance is an important quality indicator for deep groove ball bearings and cylindrical roller bearings. Given a batch of components, including outer rings, inner rings and rolling elements, with different sizes due to the manufacturing error, selective assembly is critical to properly combine these components to obtain qualified bearings as many as possible. This work proposes a method based on pointer network to address the precise selective assembly of bearing components. A graph embedding layer is employed to model all the components as graph nodes and their possible relationship as edges. The pointer network is trained to generate a sequence representing the assembly relationship of components, aiming to maximize the assembly rate and minimize the bias with respect to the best clearance. Compared with traditional intelligent optimization methods, such as genetic algorithm, which can also be used to approximate solutions for the selective assembly of bearing components, the proposed method outperforms with better generalization capability, it can be trained using a small-batch of data and be applied for a much large batch problem without losing accuracy and significant increasing the computational time.