A method, apparatus and computer program product are provided to predict faults. Initially, a plurality of features are provided to a plurality of models. A subset of features is selected for each model. The plurality of features selected by a respective model is dependent upon a time scale associated with a respective model. As a result of their dependence upon different time scales, the plurality of selected features provided to a first model will differ from those provided to a second model. The plurality of models process the respective plurality of selected features. The outputs from the plurality of models are fused to generate a measure indicative of an impending fault. By providing different selected features to the models that are dependent upon the associated time scales and by then combining the outputs of the plurality of models, the resulting measure of an impending fault may accurately predict a fault well in advance of its occurrence.