Neural Network Classifiers to Grade Parts Based on Surface Defects with Spatial Dependencies

Abstract

In many manufacturing situations, production parts must be assigned a qualitative grade, rather than only accepted or rejected. When this is done, spatial relationships between defect areas can be a critical factor in making grade assignments. In the case of grading hardwood lumber, for instance, there exists a highly complex set of grading rules which incorporate spatial information to estimate the percentage of clear wood present on a piece of lumber, and hence its grade. Algorithmic implementations of these rules are computationally time-consuming and are not guaranteed to assign a grade in a real-time manufacturing environment. This report describes an effort aimed at developing a simplified, lumber grade classifier that incorporates measures of spatial dependency among the defects. A number of different neural network classifiers are compared. Each classifier contains input features that measure contagion among the defects, as well as more traditional, non-spatial features. In contrast to the algorithmic approach, trained neural net classifiers execute in nearly constant time, i.e. effectively independent of board size, actual board grade, and number of defects. The classifier approach trades some accuracy for improved usability and operational speed. A confusion-utility matrix indicates where this trade-off is justified.

Citation

Schmoldt, Daniel L. 1995. Neural Network Classifiers to Grade Parts Based on Surface Defects with Spatial Dependencies. Review of Progress in Quantitative Nondestructive Evaluation 14A. pp. 795-802.