The data-driven deep learning methods have achieved excellent performance in the mechanical fault diagnosis of high-voltage circuit breakers. However, the premise of these methods to achieve excellent performance is the ability to obtain massive training samples. The diagnostic performance is severely degraded in scenarios where field data are scarce. To address this issue, this article proposes a novel feature fusion metric learning model for mechanical fault diagnosis of field high-voltage circuit breakers. First, a feature fusion convolutional neural network is established, which effectively improves the ability to extract discriminative features. Then, the K-nearest neighbor algorithm is used as the metric learner to realize the matching and classification of few-shot. Finally, the discriminative ability of feature representation is further improved by improving the center loss. And the transferable knowledge is learned from the large sample set constructed in the laboratory through episodic training. Experimental results show that the proposed method can achieve the diagnosis accuracy of 94. 58% when the number of samples in each type of support set is 5. It is 63. 71% higher than that of the convolutional neural network. In addition, the proposed method benefits from the episodic training method, which effectively avoids the problem of unbalanced samples.