A comparative study on the classification of engineering surfaces with dimension reduction and coefficient shrinkage methods

As more automated and accurate surface inspection devices enter the manufacturing process, engineers collect a larger amount of surface inspection data, in terms of storage space and the number of parameters to characterize the surface, but sometimes smaller in terms of the number of coherent surface observations. In these cases, more features are preferable to characterize engineering surfaces for capturing the details of the surface finish patterns. When the number of surface parameters exceeds the number of collected surface observations, a difficulty with the dimensionality emerges in classification. This paper has researched the accuracy and interpretability of using the dimension reduction and coefficient shrinkage methods in combination with the logistic model to deal with this dimensionality problem in engineering surface classification. Five methods for dimension reduction and coefficient shrinkage are selected and compared. These are: subset selection (Sub), principal component analysis (PCA), partial least squares (PLS), ridge regression (Ridge), and least absolute and shrinkage and selection operator (Lasso). A case study is used to illustrate their effectiveness by classifying 30 pump body surfaces with 40 surface feature parameters. The obtained results show that the dimension reduction methods, PCA and PLS, could achieve higher classification accuracies but their results are not interpretable. Sub could achieve higher accuracy in this case, but the discrete parameter selection process is aggressive. Finally, the classification results of the coefficient shrinkage methods, Ridge and Lasso, are interpretable for process faults diagnosis purposes; however, the accuracies are lower than the other methods.