Open-Switch and Current Sensor Faults Diagnosis in Three-Level NPC Inverters Under Small Samples

This article proposes a novel simultaneous diagnosis method for open-switch (OS) and current sensor (CS) faults in three-level neutral point clamped (NPC) inverters under small sample conditions. First, a double-discriminator auxiliary classifier generative adversarial network (DACGAN) is designed, which includes the time-domain discriminator and the time-frequency domain discriminator. Meanwhile, the model uses a wavelet-like transformer (WLT) that extracts time-frequency features of multiple faults as the input for the time-frequency discriminator. Thereby, DACGAN enables the simultaneous fitting of the data distributions in both the time domain and the time-frequency domain. Moreover, to prevent mode collapse and ensure the consistency of the temporal structure of the generated signals, DACGAN introduces Wasserstein distance, gradient penalty, and similarity loss to design a new loss function, improving the quality of the generated signals. Furthermore, a lightweight residual convolutional network incorporating efficient channel attention (ECA) modules, ECA-RLNet, is developed. Specially, to reduce the number of parameters, this network employs depthwise separable convolution and global average pooling, while residual blocks and ECA modules are introduced to enhance feature extraction and diagnostic accuracy. Finally, a hardware-in-the-loop (HIL) platform is constructed to collect multicondition data, validating the effectiveness of this method.