Roller leveler monitoring using acceleration measurements and models for steel strip properties

Abstract

The advanced steel grades and high productivity requirements in the modern steel industry subject production machines to increased mechanical stresses, which inflicts losses. Novel data-oriented solutions to the monitoring of machines have a pivotal role in loss prevention, but the industrial data with high sampling rates, noise, and dimensions bring challenges there. This study proposes a new monitoring approach for roller levelers based on vibration measurements and regression models for estimating steel strip properties including yield strength, width, and thickness. The regression residuals are monitored based on moving mean and range charts, which reveal changes from the expected normal operation. A high-dimensional feature set of 144,000 statistical features was studied with various feature selection methods, including filters and wrappers. Multiple linear regression and generalized regression neural network were applied in modeling. The approach was validated using data from an industrial roller leveler processing steel strips with diverse properties. The results reveal that the accurate prediction of the strip thickness from the strip properties is possible and multiple linear regression was generally the superior model therein. Additional simulations indicated that the control charts can detect deviant operation. Supplemental information about the momentary operation of the machine would improve the approach.