Resumen

Cascade transformer multilevel inverters (CT-MLI) are DC–AC converters used in medium and high power applications to provide standardized AC output. Despite their numerous advantages and robustness, these devices are highly susceptible to fault events because of their high amount of components. Therefore, early failure detection enables turning off the power system avoiding the propagation of the fault to the connected loads. Beyond that, converter operation can be reconfigured to tolerate the fault and activate a fail flag facilitating the subsequent corrective maintenance. The techniques proposed so far required several sensors, which is not practical. Therefore, in this study, we propose an automatic fault detection algorithm for cascade multilevel inverters based on pattern recognition, that only requires a sensor located at the output of the inverter. Naive Bayes, decision tree, nearest neighbor, and support vector machine were tested as classifiers using cross validation. The proposed method showed high detection accuracy when all the obtained descriptors were employed, being the K-NN the classifier showing superior performance. Furthermore, an evaluation was developed to determine the minimum number of descriptors required for the effective operation of the detection system, reducing the computational cost and simplifying its implementation. The method was validated by using simulation results obtained from a multilevel inverter circuit model.