A Novel Convolutional–Transformer Hybrid Neural Network Framework for Switch Fault Diagnosis in Power Converters

This paper evaluates a one-dimensional Convolutional– Transformer Neural Network (1-D CTNN) for fault diagnosis and switch-level localization in Cascaded H-Bridge (CHB) converters. The proposed architecture is validated on both 3-level and 7-level CHB topologies using experimental datasets. Different positional encoding strategies are investigated to assess their impact on diagnostic accuracy, robustness and computational efficiency. The results demonstrate consistently high performance across all tasks, including 2-class system health detection, phase and module classification, and fine-grained switch localization involving up to 37 fault classes, with accuracy approaching or exceeding 98–99%. Precision, recall and F1-scores remain comparably high, while inference times stay within the millisecond range and the model footprint remains compact, enabling realtime embedded implementation. By combining convolutional feature extraction with Transformer-based attention, the proposed 1-D CTNN effectively captures both local waveform features and long-range temporal dependencies, outperforming classical machine learning approaches and shallower neural networks. Although not intended as a universal diagnostic framework, the demonstrated scalability and efficiency indicate strong potential for extension to more complex converter architectures.