Big physics experiments can collect Tbytes (even Pbytes) of data under continuous or long pulse basis. The measurement systems that follow the temporal evolution of physical quantities translate their observations into very large time-series data and video-movies. The off-line analysis of massive amounts of waveforms and video-images is typically affected by two important limitations. Firstly, there are no methods for automatic selection to focus the analysis just on the data of interest for a particular study. Secondly, some physical events take place in a random and very infrequently way and, again, there are not automatic mechanisms to identify and locate these singular events. This article describes a universal and automatic technique to recognize and locate inside waveforms and video-films both signal segments with data of potential interest for specific investigations and singular events. The method is based on regression estimations of the signals using Support Vector Machines. A reduced number of the samples are shown as outliers in the regression process and these samples allow the identification of both special signatures and singular points. Results are given with the database of the JET fusion device.