This work proposes an innovative approach to detect, segment and classify voltage sags according to their causes. To detect and segment, Independent Component Analysis is used, with the advantage of being fast and with low computational effort in the operational stage, once it uses only 1/8 cycle of the fundamental component. For classification purposes, Higher-Order Statistics are used for feature extraction and the classifiers are based on Neural Networks and Support Vector Machines. It was tested signal windows of 1, 1/2, 1/4 and 1/8 cycle. For both detection/segmentation design and feature selection, it was used the metaheuristics Teaching-Learning-Based Optimization. Encouraging results were achieved for the simulated signals. In addition, real signals were used to evaluate the detection and segmentation method and good results were achieved in which a detection error rate of 0.86% was achieved.