Statistical tolerance allocation design considering form errors based on rigid assembly simulation and deep Q-network

• Published in: International journal of advanced manufacturing technology 2020-12, Vol.111 (11-12), p.3029-3045
• Main Authors: He, Ci, Zhang, Shuyou, Qiu, Lemiao, Wang, Zili, Wang, Yang, Liu, Xiaojian
• Format: Article
• Language: English
• Subjects: Assembly; CAE) and Design; Computer-Aided Engineering (CAD; Engineering; Industrial and Production Engineering; Jacobi matrix method; Jacobian matrix; Machining; Mechanical Engineering; Media Management; Monte Carlo simulation; Original Article; Particle swarm optimization; Production costs; Statistical methods
• ISSN: 0268-3768; 1433-3015
• DOI: 10.1007/s00170-020-06283-w

Consideration of form errors involves real machining features in tolerance modeling but increases uncertainties in functional requirement estimation, when tackling the trade-off between the cost and precision performance. In this paper, a statistical tolerance allocation method is presented to solve this problem. First of all, a top-down stepwise designing procedure is designed for complex products, and a combination of Jacobian matrix and Skin Model Shapes is applied in modeling the mechanical joints. Then, rigid assembly simulations of point-based surfaces are further advanced to provide an accurate estimation of the assembly state, through considering physical constraints and termination conditions. A mini-batch gradient descent method and a backtracking strategy are also proposed to promote computational efficiency. Finally, a deep Q-network is implemented in optimal computation after characterizing the systematic state, action domain, and reward function. The general tolerance scheme is then achieved using the trained Q-network. The results of 6 experiments each with 200 samples show the proposed method is capable of assessing tolerance schemes with 35.2% and 47.2% lower manufacturing costs and 16.7% and 28.3% higher precision maintenance on average than conventional particle swarm optimization and Monte Carlo method respectively.