Positional Embedding Comparison for Improved 1-D Convolutional Transformer Hybrid Neural Network Fault Diagnosis on Power Converters

A comparative analysis on different positional encoders in a 1-D Convolution Transformer Hybrid Neural Network for Open Switch Fault Detection and Localization in a multilevel Cascaded H-Bridge (CHB) Converter is presented in this paper. Specifically, the performance efficacy evaluation of the sinusoidal, rotational and learnable positional encoders are compared, demonstrating their performance on multi-class classification models for open-circuit switch fault detection and localization. The models are applied for detecting the overall system health status, localizing faulty phase modules and single switch faults, respectively achieving accuracy values reaching 99% in each case of the three-level CHB converter. Additionally, Fourier positional encoder, T5 positional encoder and Relative positional encoder were considered in the seven-level CHB converter which in each cases the accuracy is above 95%.