This paper describes how a condition-based maintenance plan was developed for a combined-cycle power plant at a medium-sized Italian refinery. Including forecasting activities in the maintenance cycle achieved the dual goal of identifying any need for measures ahead of the deadlines established for routine preventive maintenance in the event of alarm conditions being detected, and of postponing any scheduled measures in the event of the components in question still being in good condition. Soft computing tools were experimentally used to achieve these objectives. Recurrent neural nets and neuro-fuzzy systems were used to ensure that the assessment of the trends of the global values was effective in determining the time remaining before the next outage period was needed. Using these tools enabled an accurate prediction of the values of the vibrations on rotating machinery based on the values of the operating parameters given as input. The plan was part of a maintenance management scheme seen as a container of inspection activities providing the foundations for systematically organizing certain servicing measures (e.g. the replacement of bearings, or alignments on rotating machinery, etc.), and to prevent sudden breakdown situations. (c) 2005 Elsevier Ltd. All rights reserved.