Aiming at the problem that the recognition accuracy of the model is not high due to the complex and variable engineering environment， a rolling bearing fault diagnosis model integrating Markov transition field and graph attention networks （MTF GAT） is proposed in this paper. Using the advantage of MTF to retain the time correlation of the signal is applied to transform one dimensional signals into two-dimensional feature maps， and the nodes and edges of the graph are defined. The graph attention layer can adaptively assign different weights to adjacent nodes to improve the ability of the model to capture useful fault features， and the abstract information of the graph is further extracted through the deep convolution module. By simulating the actual engineering en? vironment， the various fault signals are input into the trained MTF-GAT model for fault diagnosis， and the model is verified by ex? periments on two data sets. The results show that the proposed model in this paper can accurately complete the task of fault classifi? cation in a variety of environments. Compared with other deep learning models， the MTF-GAT model has better recognition accu? racy and generalization performance.