Tensile strength prediction of steel sheets: an insight into data-driven models, dimensionality reduction, and feature importance

Abstract In this work we apply data-driven models for predicting tensile strength of steel coils from chemical composition and process parameters. The data originates from steel production and includes a full chemical analysis, as well as many process parameters and the resulting strength properties from tensile tests. We establish a data pre-processing pipeline, where we apply data cleaning and feature engineering to create a machine-readable dataset suitable for various modeling tasks. We compare prediction quality, complexity and interpretability of pure machine learning (ML) models, either with the full feature set or a reduced one. Dimensionality reduction methods are used to reduce the number of features and therefore reduce complexity, either with a smart selection method or feature encoding, where features are combined and the included information is preserved. In order to determine key features of our models, we are investigating feature importance ratings, which can be used as a feature selection criteria. Furthermore, we are highlighting methods to explain predictions and determine the impact of every feature in every observation applicable for any ML model.