Power transformers' failures carry great costs to electric companies since they need resources to recover from them and to perform periodical maintenance. To avoid this problem in four working 40 MVA transformers, the authors have implemented the measurement system of a failure prediction tool, that is the basis of a predictive maintenance infrastructure. The prediction models obtain their inputs from sensors, whose values must be previously conditioned, sampled and filtered, since the forecasting algorithms need clean data to work properly. Applying data warehouse (DW) techniques, the models have been provided with an abstraction of sensors the authors have called virtual card (VC). By means of these virtual devices, models have access to clean data, both fresh and historic, from the set of sensors they need. Besides, several characteristics of the data flow coming from the VC, such as the sample rate or the set of sensors itself, can be dynamically reconfigured. A replication scheme was implemented to allow the distribution of demanding processing tasks and the remote management of the prediction applications.