A NEURAL NETWORK BASED APPROACH FOR TOLERANCE ANALYSIS

Tolerance analysis consists of determining upper bounds on the variations in a product's part dimensions such that after the parts are assembled the accumulated effect of these variations will not affect the product quality. In the past, tolerance analysis focused on the design stage in a mass production environment where the distribution of parts and components was necessarily known in advance. Thus, the assumptions concerning statistical distributions found in previous research on tolerance analysis are not applicable in flexible manufacturing and assembly environments. This research focuses on tolerance analysis for low volume, large variety production, in which parts with certain characteristics in common are typically interchangeable, as in the modular flexible assembly environment. We apply neural network techniques to predict assembly tolerances without a priori assumptions concerning the statistical distributions of parts and components. Experiments have shown that the proposed neural network based approach outperforms the traditional tolerance analysis model when the distributions of part tolerances are uniform, Weibull or mixed.