This paper describes the development of a new 'cascade' monitoring system for the aluminium smelting process that uses latent variable models. This system is based on the changes of variability patterns within a feeding cycle which are used to provide indications of faults and their possible causes. The system has been tested offline using 31 data sets. The performance of the system to detect an anode effect has been compared with a typical latent variable model that monitors the change of behaviour at every time instant. The results show that the 'cascade' monitoring system is able to detect abnormal events. It was possible to relate each event with specific patterns associated with abnormalities thus facilitating later fault diagnosis. (C) 2011 Elsevier Ltd. All rights reserved.