Robust Open-Switch Fault Diagnosis of Three-Level NPC Inverters Based on Data Augmentation With White Noise Injection
Ji-Won Jung, Dyan Puspita Apsari, Dong‐Choon Lee
Abstract
This article proposes a novel real-time fault diagnosis approach for three-level neutral-point-clamped inverters based on a one-dimensional (1-D) convolutional neural network (CNN). The proposed method incorporates data augmentation into simulation data, enhancing the generalization capabilities of deep learning models. This allows fault diagnostic models to have high robustness even in untrained system conditions. In such scenarios, the application of 1-D CNN models with data augmentation surpasses the performance of the same models without the incorporation of white noise, resulting in accuracy improvements of up to 1.71%. Furthermore, deep learning models trained on simulation data with data augmentation give a better performance when compared to those trained using experiment data. The proposed method has been verified through simulation with offline testing and experimentation with real-time deep learning algorithms.