# Fault Detection of T-Type Three-Level Converters with Simulation-Data Transfer Learning Strategy

## TL;DR

This paper introduces a new fault detection framework for multilevel converters using simulation data and transfer learning, reducing the need for real-world fault data.

## Contribution

The novel STLNet framework uses simulation data and transfer learning to improve fault detection in converters with limited real data.

## Key findings

- STLNet achieves superior diagnostic accuracy compared to traditional methods.
- The framework reduces dependency on real-world fault data through simulation-based pre-training.
- Symmetry-based augmentation enriches fault samples for better model training.

## Abstract

Accurately locating switching device faults in multilevel converters remains a challenge, particularly considering the scarcity of labeled fault data in practical industrial applications. To address this, this paper proposes a data-driven fault detection framework based on a simulation transfer learning network (STLNet). First, raw three-phase current signals are preprocessed using resampling, wavelet denoising, and normalization to generate 2D current feature images. To enrich the fault samples, a symmetry-based augmentation strategy is applied. Subsequently, a lightweight convolutional neural network is pre-trained on abundant simulation data to learn fundamental fault signatures. Finally, the designed model is transferred to the real domain by fine-tuning with a minimal amount of experimental data. Experimental validation on a T-type three-level converter platform demonstrates that the proposed STLNet achieves superior diagnostic accuracy and generalization performance compared with traditional methods, while significantly reducing the dependency on real-world fault data.