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

Area of Science:

• Electrical Engineering
• Power Electronics
• Artificial Intelligence

Background:

• Accurate fault location in multilevel converters is crucial but challenging due to limited labeled fault data.
• Existing methods struggle with the scarcity of real-world fault datasets for training.

Purpose of the Study:

• To propose a data-driven fault detection framework, the simulation transfer learning network (STLNet), to overcome data scarcity.
• To improve the accuracy and generalization of fault diagnosis in multilevel converters.

Main Methods:

• Preprocessing three-phase current signals into 2D feature images using resampling, wavelet denoising, and normalization.
• Employing a symmetry-based augmentation strategy to increase fault sample size.
• Pre-training a lightweight convolutional neural network on simulation data and fine-tuning with minimal experimental data.

Main Results:

• The proposed STLNet achieved superior diagnostic accuracy and generalization performance on a T-type three-level converter.
• The framework significantly reduced the dependency on extensive real-world fault data.
• Demonstrated effectiveness compared to traditional fault detection methods.

Conclusions:

• The STLNet framework offers an effective solution for fault detection in multilevel converters with limited data.
• Simulation transfer learning is a viable approach to enhance diagnostic capabilities in industrial applications.
• The method provides a practical and efficient alternative for ensuring converter reliability.