Densely-connected Decoder Transformer for Unsupervised Anomaly Detection of IGBT System

Detecting anomalies in IGBT-based power systems is crucial for ensuring safety. With superior performance in various fields, transformer networks are promising for anomaly monitoring. However, with the structure designed for different tasks, certain adaptation is needed for better detection results. Another issue of data-driven methods like transformer networks is the anomalous data required for training, which may be difficult to collect in practice. Therefore, a novel model, Densely- connected Decoder Transformer (DDformer), is proposed for unsupervised anomaly detection of IGBT systems in this paper. A transformer-like decoder with dense connections and no recursion is developed, combining features from all encoder layers. Then, a platform of three-phase full-bridge inverter is built. Finally, an experiment is conducted on the platform to demonstrate the superiority of DDformer.