The complete training data under target working conditions are necessary for the traditional data-driven fault diagnosis methods of machinery. However, the actual working conditions of mechanical equipment are complicated and difficult to predict, and thus it is difficult to obtain sufficient training data. To solve this problem, this paper proposed a deep embedding metric network (DEMN) for mechanical fault identification across different working conditions. The proposed method uses the data under known working conditions to learn robust feature representation, and then establish the generalized fault diagnosis model for the unseen working conditions. First, the deep embedding features of fault signal are extracted by the multiscale convolutional neural network (MCNN). Then, the triplet loss-based metric learning objective is optimized to enhance the discriminant ability of classification boundary. Particle swarm optimization (PSO) algorithm is executed to the search the optimal margin in triplet loss. By facilitating the intra-class compactness and the inter-class separability, the influence of working condition changes to fault relationship mapping is significantly reduced. The experimental results show that the proposed method presents superior accuracy and generalization performance in gearbox fault diagnosis across different working conditions.