Deep Feature Selection Framework for Quality Prediction in Injection Molding Process

In the injection molding process, accurate quality prediction of manufactured products is vital for process optimization. With the development of machine learning, several studies have been conducted to predict product quality, but practically many variables used in injection molding compared with generatable process samples cause the curse of dimensionality. We propose the novel deep network-based feature selection method to solve the overfitting problem and improve quality prediction for various geometrical shapes and resin materials. The proposed method introduces a layerwise relevance propagation for the regression task (LRP-R) to measure the feature importance in the regression task. In addition, recursive feature elimination with cross-validation generates an optimal subset by selecting features based on the feature importance measured in LRP-R. Finally, the proposed method is used to interpret feature importance and improve quality prediction in the injection molding process. Comparative experiments show the proposed method's statistical validity and domain knowledge-based validation.