Edge Localized Modes (ELMs) are instabilities occurring in the edge of H-mode plasmas. Considerable efforts are being devoted to understanding the physics behind this non-linear phenomenon. A first characterization of ELMs is usually their identification as Type I or Type III. An automated pattern recognition system has been developed in JET for off-line ELM recognition and classification. The empirical method presented in this paper analyzes each individual ELM instead of starting from a temporal segment containing many ELM bursts. The ELM recognition and isolation is carried out using three signals: Da, line integrated electron density and stored diamagnetic energy. A reduced set of characteristics (such as diamagnetic energy drop, ELM period or Da shape) has been extracted to build supervised and unsupervised learning systems for classification purposes. The former are based on Support Vector Machines (SVM). The latter have been developed with hierarchical and k-means clustering methods. The success rate of the classification systems is about 98% for a database of almost 300 ELMs.