This paper presents an original research initiated by the monitoring needs of a semiconductor production plant. The industrial operations rely on an Information Technology (IT) system, and several time series data are controlled statistically. Unfortunately, these variables often contain outliers, as well as structural changes due to external decisions in the IT activity. As a consequence, it has been observed that the monitoring results obtained with standard techniques could be severely biased. This paper presents some contributions to overcome such difficulties. A new monitoring method is proposed, based on robust Holt-Winters smoothing algorithm, and coupled with a relearning procedure for structural breaks detection. Such a method is flexible enough for a large-scale industrial application. We evaluate its performances through simulations studies, and show its usefulness in industrial real applications for univariate and multivariate time series. The scope of application deals with IT activity monitoring, but the introduced statistical methods are generic enough for being used in other industrial fields.