Semantic-guided feature selection for industrial automation systems

Abstract

Modern industrial automation systems incorporate a variety of interconnected sensors and actuators that contribute to the generation of vast amounts of data. Although valuable insights for plant operators and engineers can be gained from such data sets, they often remain undiscovered due to the problem of applying machine learning algorithms in high-dimensional feature spaces. Feature selection is concerned with obtaining subsets of the original data, e.g. by eliminating highly correlated features, in order to speed up processing time and increase model performance with less inclination to over-fitting. In terms of high-dimensional data produced by automation systems, lots of dependencies between sensor measurements are already known to domain experts. By providing access to semantic data models for industrial data acquisition systems, we enable the explicit incorporation of such domain knowledge. In contrast to conventional techniques, this semantic feature selection approach can be carried out without looking at the actual data and facilitates an intuitive understanding of the learned models. In this paper we introduce two semantic-guided feature selection approaches for different data scenarios in industrial automation systems. We evaluate both approaches in a manufacturing use case and show competitive or even superior performance compared to conventional techniques.