Estimating Distributions of Surface Parameters for Classification Purposes

Abstract

The classification of workpiece surface patterns is an essential element in trying to understand how functional performance is influenced by the surface geometry. Filter banks have been investigated in literature for capturing the multiscale characterization of the engineering surfaces. Conventionally, parametric representations of the filter outputs were used for classification. In this paper, the histogram estimators of the filter bank outputs from engineering surfaces in combination with the nearest neighbor method for classification are investigated to improve the classification accuracy, which are accomplished by utilizing distribution dissimilarity measures to compare histograms. Furthermore, for large and complex surfaces, the histogram estimators of local surface flatness parameters are also proposed for the purpose of simple computation. Two case studies have been conducted to demonstrate the proposed methods. Influence of the histogram bins for histograms and the dissimilarity measures on classification performance is studied in detail. Results from the first case study show that the proposed method is less effective in classifying small surfaces with clear surface patterns, because the filtering is influenced by the quality of the surface data collected from the measurement sensor. In comparison, results from the second case study show that the proposed method performs better in classifying large surfaces with mild surface pattern differences. The classification accuracy using the conventional method drops from 100% to around 50% in the second case study. In general, one can achieve misclassification errors below 5% in both case studies with the histogram representations of surface parameters and the appropriate selection of the number of bins for histogram construction.