The effect of feature reduction and feature selection on prediction the characteristics of the outputs in multistage manufacturing

Abstract

In a complex manufacturing system such as the multistage manufacturing system, maintaining the quality of the products becomes a challenging task. This is due to the interconnectivity and dependency of factors that can affect the final product. With the increase availability of data, machine learning approaches are applied to assess and predict quality-related issues. In this paper, several machine learning algorithms including feature reduction/selection methods were applied to a publicly available multistage manufacturing dataset to predict the characteristic of the output measurements in (mm). A total of 24 prediction models was produced. The accuracy of the prediction models and the execution time were the evaluation metrics. The results show that uncontrolled variables are the most common features that have been selected by the selection/ reduction methods suggesting their strong relationship to the quality of the product. Also, some positive and negative impact on the accuracy and execution time has been observed after the application of feature selection/ reduction. However, the performance of the prediction models was heavily dependent on the machine learning algorithm.