PAS - A Feature Selection Process Definition for Industrial Settings

Abstract

Industry 4.0, and the accompanying digitalization of industrial processes, lead to a continuously increasing amount of data. One reason is the installation of more sensors on machines and production lines in combination with shorter measurement intervals. The rising amount of data requires elaborate methods to remove redundant and uninformative data to reduce time complexity and required computational resources. This could be achieved by feature selection methods in combination with the right preprocessing and enhancement methods to generate and identify important features. Feature selection methods are mostly described in a methodological way without knowing how data is handled prior to their application, assuming a high-quality dataset. To fill this gap, we define a process for feature selection, focusing on time-series industrial data. The definition consists of three phases for the data adaptation: Preprocessing, Amplification and Selection. Each phase fulfills a specific purpose, ranging from handling missing and censored data to the enhancement of the dataset by higher-order data and aggregations and the selection itself. Our experiments show that the application of these phases on three different industrial use-cases lead to better accuracy results in machine learning models.