Prediction of Product Quality in Continuous Glass Manufacturing Process
Pedro Pereira Rodrigues, João Manuel Portela da Gama · 2004
This paper presents a study made in the scope of the EUNITE 2003 Competition. We have designed a learning system with four main pre-processing steps. First we smooth data variables to reduce possible high-frequency noise, separating each time series in trend, seasonality, and noise; the goal here is to model trend. After this, we cluster input variables using a divisive clustering algorithm. The clusters’ centroids are used as input variables to the system. The system then models the firstorder differences, which is an important factor to the fourth aspect of the system: analysis of changes in the distribution of the examples. This analysis was made either using a concept drift detection method based on the error of on-line predictions or monitoring the residuals of a fitted linear model. These methodologies were tested with real-world data sets from glass manufacturing industry. The experimental results show few evidences of a need for training period selection, but a good performance at the change of concept detection and also with learning the new concept. Residuals monitoring technique proved to be a simpler but quicker and competitive process when compared with concept drift detection.