Connectors are essential components of electronic devices, and the cleanliness of their working contacts is a necessary condition for the normal operation of electronic equipment. To address the frequent issues of false positives and false negatives caused by the diverse types and styles of connectors, as well as the limited and variable foreign object samples, this article proposes a novel zeroshot anomaly detection method. By synthesizing random anomalies on unrelated background images, it constructs pairs of normal-anomaly sample images. Through network prediction, a discrepancy score map representing the pixel-level similarity between sample pairs is obtained, enabling anomaly detection and localization. By employing anomaly region mask supervision, the network focuses on the pixel differences between normal and anomaly samples, reducing the network′s attention to the semantic information of the images themselves and minimizing the need for real samples. Thus, the generalization ability of the detector is enhanced. To evaluate the effectiveness of the algorithm, the network is trained solely on synthesized data and evaluated on the DeepPCB dataset, achieving a mAP (mean average precision) of 88. 2% . After transfer learning, the mAP increases to 99. 1% , which is the best performance on this dataset to date. Experimental results demonstrate the strong generalization ability of the proposed zero-shot anomaly detection method.