Semi-automatic Maintenance of Regression Models: an Application in the Steel Industry

Abstract

Software applications used in the controlling and planning of production processes commonly make use of predictive statistical models. Changes in the process involve a more or less regular need for updating the prediction models on which the operational software applications are based. The objective of this article is
- to provide information which helps to design semiautomatic systems for the maintenance of statistical prediction models and
- to describe a proof-of-concept implementation in an industrial application.
The system developed processes the production data and provides an easy-to-use interface to construct updated models and introduce them into a software application. The article presents the architecture of the maintenance system, with a description of the algorithms that cause the system’s functionality. The system developed was implemented for keeping up-to-date prediction models which are in everyday use in a steel plate mill in the planning of the mechanical properties of steel products. The conclusion of the results is that the semi-automatic approach proposed is competitive with fully automatic and manual approaches. The benefits include good prediction accuracy and decreased workload of the deployment of updated model versions.