A learning procedure for detection of process anomalies in the production of metal long products and a new industrial case study

The highly individualized production processes for long products in the steel industry is subject to a variety of influencing variables with mutual interactions in a complex manner. To handle this complexity, modern data mining methods can be used for a highly efficient analysis of process data, to detect process anomalies in the process data, e.g. from rolling mills by statistical pattern recognition. This paper proposes a data-based strategy for detecting process anomalies within a hot rolling mill for long products. Suitable data are identified and selected from existing sensors and processed within a new database. This central database is used to train classification algorithms. The reliability of two prominent classifiers based on Principal Component Analysis (PCA) and One-Class Support Vector Machines (OC-SVM) has been evaluated. From the comparison in this respective use case, it has been concluded that satisfying results can be obtained, but PCA is highly dependent on the data distribution. The OC-SVM has also been implemented and tested and offers advantages when the data sets have a more complex distribution.