Feature selection techniques and comparative studies for large-scale manufacturing processes

Abstract

In order to efficiently and effectively control an overall process in the process industry, a few important parameters should be identified from high-dimensional, non-linear, and correlated data. Feature selection techniques can be employed to extract a subset of process parameters relevant to product quality. The performance of these techniques depends on the precision of the prediction model formulated to quantify the relationship between the process parameters and the quality characteristics. Although the neural network-based partial least squares (NNPLS) method has been proven to be effective in prediction models for the aforementioned industrial process data, feature selection techniques appropriate for NNPLS models have yet to appear. Here, several techniques for scoring the relevance of process parameters to product quality are proposed and validated by applying three datasets. These experiments show that the proposed techniques can discriminate relevant process parameters from irrelevant ones.