Enhanced Open-Circuit Fault Diagnosis in T-Type Inverters Using Conditional Virtual Sample Generation

In real-world electric power systems, electrical faults often result in significant economic losses, making it critical to prevent their occurrence whenever possible. Consequently, fault samples are typically scarce and challenging to acquire. To address this issue, we propose a novel generative model that enables the controlled generation of large quantities of virtual fault samples. Unlike existing conditional Wasserstein generative adversarial networks (CWGANs), we introduce a novel conditional batch normalization (CBN) module. This module jointly normalizes the input noise distribution and label embeddings, ensuring stable training and precise control over the generated outputs. Furthermore, this design mitigates interclass mode collapse, which is a common issue in conditional generation caused by significant disparities in label embeddings. In addition, we propose a temporal convolutional network enhanced with squeeze-and-excitation (TCN-SE) component equipped with SE blocks for adaptive feature recalibration, which effectively captures temporal dependencies in the fault data and performs accurate fault classification. Finally, detailed experiments are conducted to evaluate the effectiveness of the proposed model. The results demonstrate that the proposed generative module achieves an average improvement of 4.46% across six different models. Additionally, the proposed TCN-SE model achieves the highest accuracy of 96.36%. This study provides a robust framework for fault data augmentation and diagnosis in power electronic systems, contributing to the reliability and maintenance of three-phase T-type inverters.