Three-phase pulse width modulation converters using insulated gate bipolar transistors (IGBTs) have been widely used in industrial application. However, faults in IGBTs can severely affect the operation and safety of the power electronics equipment and loads. For ensuring system reliability, it is necessary to accurately detect IGBT faults accurately as soon as their occurrences. This paper proposes a diagnosis method based on data-driven theory. A novel randomized learning technology, namely extreme learning machine (ELM) is adopted into historical data learning. Ensemble classifier structure is used to improve diagnostic accuracy. Finally, time window is defined to illustrate the relevance between diagnostic accuracy and data sampling time. By this mean, an appropriate time window is achieved to guarantee a high accuracy with relatively short decision time. Compared to other traditional methods, ELM has a better classification performance. Simulation tests validate the proposed ELM ensemble diagnostic performance.