This paper presents a data driven approach for failure prediction for Electrical Submersible Pumps (ESP). ESP system is well known as an effective artificial lift method which has been applied to about 20 percent of almost one million wells worldwide. Well failures lead to production loss and generally the repair cost for an ESP is usually much higher than those of other artificial lift systems, thus predicting ESP failures before they occur will be valuable. We apply advanced machine learning techniques for predicting ESP failures using electrical and frequency data from the field. Data from real-world assets using ESP systems is analyzed to learn examples of normal well and failure conditions. A generalized Support Vector Machine (SVM) trained with a set of selected features is developed and this approach is tested on real world data. Our results show that this approach works well based on feedback from subject matter experts on the results.