Abstract: To address the challenges posed by scarce DC bias fault samples in power transformers—which lead to low accuracy and high false alarm rates in data-driven diagnostic models—this study proposes an acoustic anomaly detection method that integrates self-supervised pre-training and metric learning. First, we conduct an in-depth analysis of the acoustic mechanisms underlying core magnetostriction and winding vibration forces under DC bias conditions. Second, we develop a self-supervised learning framework by designing proxy tasks with transformer voltage levels and environmental interference types as pseudo-labels, enabling the Conformer encoder to learn highly discriminative acoustic features using only normal operational data. Finally, the K-nearest neighbor algorithm computes distances between test samples and the normal feature manifold to identify anomalies. Experiments on real-world datasets comprising transformers operating at multiple voltage levels demonstrate that this method achieves an AUC of 99.84% and an F1 score of 99.01% in DC bias fault detection. Compared with standard Transformers and various unsupervised models, this approach effectively captures subtle fault characteristics, providing a novel solution for industrial anomaly detection under zero-shot (i.e., no-fault-labeling) conditions.