Fault diagnosis method of a cascaded H‐bridge inverter based on a multisource adaptive fusion CNN‐transformer

Abstract

AbstractIn high‐voltage applications, the number of cascaded H‐bridge inverter units is large, the failure probability increases, and the waveform similarity is high after the failure of power devices at different positions. This paper proposes a fault diagnosis method for high‐voltage multilevel cascaded H‐bridge inverter based on a multisource adaptive fusion CNN‐transformer. The method transforms the prefiltered three‐phase multilevel voltage and postfiltered three‐phase current waveforms into wavelet time‐frequency maps using a continuous wavelet transform. The time‐frequency maps of the six signal sources are used as inputs to the network. The convolutional neural network is employed to extract fault features, resulting in six feature maps. These six feature maps are then assigned certain weights to generate a new fused feature map, which is input into the transformer for training and achieving fault state recognition in cascaded H‐bridge inverters. The experimental results show that the method can effectively identify similar fault features. The diagnostic accuracy is as high as 99.15% in the measured data of the experimental platform, and the diagnostic time is 0.069 s. It has high fault identification accuracy and provides a new idea for fault diagnosis of high‐voltage multilevel cascaded H‐bridge inverters.