The traditional graph-based fault diagnosis framework usually uses a certain structural relationship of high-dimensional dataset to construct a similarity graph to reveal the geometric structure between samples, resulting in the loss of other structural information of the dataset, and it is impossible to accurately extract the low-dimensional features that characterize the running state of the bearing. A new graph-based unsupervised feature extraction method is proposed, which considers both the global and local structures of high-dimensional dataset in the process of constructing graphs, which is called global-local graph embedding. The method first constructs an undirected graph by using the global structure information of the dataset. Then, by constructing local structure information and assigning corresponding weights to the edges in the undirected graph, a global-local graph joint representation convex optimization problem is obtained, and the similarity between samples is evaluated according to the obtained weights. Finally, the low-dimensional embedding result is calculated by keeping the similarity between samples unchanged in the low-dimensional space. Compared with the single graph structure representation, our constructed global-local joint graph takes full advantage of the global and local structural information inherent in high-dimensional dataset. In addition, the essential features of high-dimensional bearing data can be effectively extracted by maintaining the similar performance between samples. Experimental results show that our proposed feature extraction method based on global-local graph embedding has obvious advantages over existing methods.