To address the problems of different data distribution of rolling bearings under different speed conditions and low fault diagnosis accuracy caused by insufficient label samples in practical engineering applications, the domain adaptation modules are integrated into masked autoencoders (MAE). An improved masked autoencoders (IMAE) method for semi-supervised fault diagnosis of rolling bearings is proposed. Firstly, the two-dimensional time-frequency graph of the response signal is obtained by applying continuous wavelet transform (CWT) to the vibration signal of the rolling bearing. Then, the mask of the time-frequency graph is randomly masked, and the mask autoencoder is pre-trained with unlabeled samples to obtain the complex intrinsic features of the data. The reliance on labeled samples is reduced. Secondly, the domain adaptation module is introduced into the pre-trained encoder, and a small amount of labeled source domain data are used to fine-tune the IMAE, and the maximum mean difference is minimized in Hilbert space to reduce the data distribution difference between the source domain and the target domain caused by different rotational speeds. Finally, the semisupervised fault diagnosis of rolling bearing is realized under the Softmax classification layer. Through the experimental evaluation of the rolling bearing data set, the detection accuracy of the proposed method is more than 94% , which proves the feasibility and effectiveness of the proposed method.