Semi-supervised roughness prediction with partly unlabeled vibration data streams

Experimental data sets that include tool settings, tool and machine-tool behavior, and surface roughness data for milling processes are usually of limited size, due mainly to the high costs of machining tests. This fact restricts the application of machine-learning techniques for surface roughness prediction in industrial settings. The primary objective of this work is to investigate the way data streams that are missing product features (i.e. unlabeled data streams) can contribute to the development of prediction models. The investigation is followed by a proposal for a semi-supervised approach to the development of roughness prediction models that can use partly unlabeled data to improve the accuracy of roughness prediction. Following this strategy, records collected during the milling process, which miss roughness measurements, but contain vibration data are used to increase the accuracy of the prediction models. The method proposed in this work is based on the selective use of such unlabelled instances, collected at tool settings that are not represented in the labeled data. This strategy, when applied properly, yields both extended training data sets and higher accuracy in the roughness prediction models that are derived from them. The scale of accuracy improvement and its statistical significance are shown in the study case of high-torque face milling of F114 steel. The semi-supervised approach proposed in this work has been used in combination with supervised kNearest Neighbours and random forest techniques. Furthermore, the study of both continuous and discretized roughness prediction, showed higher gains in accuracy in the second.