Reliable electricity infrastructure is critical for modern society, highlighting the importance of securing the stability of fundamental power electronic systems. However, as such systems frequently involve high-current and high-voltage conditions, there is a greater likelihood of failures. Consequently, anomaly detection of power electronic systems holds great significance, which is a task that properly-designed neural networks can well undertake, as proven in various scenarios. Transformer-like networks are promising for such application, yet with its structure initially designed for different tasks, features extracted by beginning layers are often lost, decreasing detection performance. Also, such data-driven methods typically require sufficient anomalous data for training, which could be difficult to obtain in practice. Therefore, to improve feature utilization while achieving efficient unsupervised learning, a novel model, Densely-connected Decoder Transformer (DDformer), is proposed for unsupervised anomaly detection of power electronic systems in this paper. First, efficient label-free training is achieved based on the concept of autoencoder with recursive-free output. An encoder–decoder structure with densely-connected decoder is then adopted, merging features from all encoder layers to avoid possible loss of mined features while reducing training difficulty. Both simulation and real-world experiments are conducted to validate the capabilities of DDformer, and the average FDR has surpassed baseline models, reaching 89.39%, 93.91%, 95.98% in different experiment setups respectively.