Abstract14,000 data sets from an industrialized bainitization process, consisting of process gas furnace, salt bath and circulating air furnace, were used to predict the resulting Vickers hardness of cylinder heads made of 100Cr6 based on process data such as temperature and pressure. For prediction, machine learning methods such as ANNs, CNNs, ensemble methods and support vector regressors were compared. Meta features such as the furnace number as well as features extracted from the recorded time series were used. Data preparation and feature extraction were performed according to the machine learning methods used. The random forest achieved the best predictions with an R2 score of 0.406 and also allows the evaluation of the most important features.

No abstract

To explain the variance in core hardness of 18CrNi8 nozzle bodies after industrial heat treatment, several data sources, including steel melt composition, sensor process data, and measurement errors, of five years are aggregated. In order to predict hardness variations caused by alloy composition, traditional physical models by Maynier are compared with data-driven machine learning models, which show no advantage due to low data variability. Neither method can fully explain the visible drifts, which are better tracked by an alternative (i. e., filter model) that uses past measurements. Machine learning on features from heat treatment is not successful in predicting hardness change, presumably because the process is too stable. Finally, a large part of the variance is caused by the HV 1 measurement error.