Rolling Average-Decision Tree-Based Fault Detection of Neutral Point Clamped Inverters

Abstract

Listen

Translate article

Multilevel inverters (MLI) with many switches, which are assets for EVs and other high and medium power applications are prone to switch faults. To ensure the reliable operation of MLI, simple, effective, and automatic fault monitoring and localization techniques are warranted. This paper focus on pole voltage rolling average (RA) based generalized open switch fault detection (FD) of neutral point clamped (NPC) MLI. The RA contributes advantage of smoothing the input, thus reducing the false alarm due to switching transients and standard deviation of averaged pole voltage serves as a good fault feature to distinguish between inner and outer switch faults for single switch fault case and the correct switch pair for double switch fault case. The proposed method based on pole voltage RA is fast and load-independent and uses decision tree tuned with random search as the classification tool for 12+1 and 78+1 classes, i.e., all possible chances of single and double switch fault combinations with training accuracy of 99% and 97.5% for three- level NPC. The paper also includes an analysis of how the DT classifier is affected by overfitting as the tree's depth increases.