Three-Phase Inverter Fault Diagnosis Based on an Improved Deep Residual Network

Abstract

This study addresses the challenges of limited fault samples, noise interference, and low accuracy in existing fault diagnosis methods for three-phase inverters under real acquisition conditions. To increase the number of samples, Wavelet Packet Decomposition (WPD) denoising and a Conditional Variational Auto-Encoder (CVAE) are used for sample enhancement based on the existing faulty samples. The resulting dataset is then normalized, pre-processed, and used to train an improved deep residual network (SE-ResNet18) fault diagnosis model with a channel attention mechanism. Results show that the augmented fault samples improve the diagnosis accuracy compared with the original samples. Furthermore, the SE-ResNet18 model achieves higher fault diagnosis accuracy with fewer iterations and faster convergence, indicating its effectiveness in accurately diagnosing inverter open-circuit faults across various sample situations.