An Optimized Rotational Position Encoding for 1-D Convolutional Transformer Hybrid Neural Network Fault Diagnosis on Power Converters

The rotational position encoder is here adopted and optimized on a 1-D Convolutional Transformer Hybrid Neural Network model for fault diagnostics on power converters. The effectiveness of the proposed algorithm is validated on two different Cascaded H-Bridge (CHB) topologies: a 3-level and a 7-level configuration, thereby testing its scalability to systems with a higher number of switches. The evaluation focuses not only on classification accuracy but also on model robustness and reduction in model size. Given the inherent complexity of power converter topologies, the results demonstrate that the proposed model effectively processes dynamic time-series signals to detect and localize open-switch faults. The performance of both binary and multi-class classifiers confirms the algorithm's capability across varying fault scenarios and converter configurations.