The remaining useful life of rolling element bearing affects the reliability and stability of the machine. Accurately predicting the remaining useful life of rolling element bearings is always necessary to make maintenance decisions in practical engineering. However, the remaining useful life prediction methods using traditional machine learning are incompetent for the task of training without labels, which consumes computational cost, financial resources, and labor. Therefore, a gated recurrent unit model based on feature dimensionality reduction and the multi-head self-attention mechanism (MHGRU) is established, and a homology algorithm is proposed to train the model and make the prediction. The isometric feature mapping algorithm and a homology algorithm are used in the model training, which incorporates the advantages of appreciable computational efficiency and preservation of automatic labeling for training in engineering. First, 24 basic features are extracted from the life-cycle vibration signals of rolling element bearings to reconstruct fusion features by the isometric feature mapping algorithm, which aims to reduce the feature dimension and improve computational efficiency. Since the multi-head self-attention mechanism in the MHGRU has the ability to comprehensively highlight the attention coefficient for long-term fusion features at different moments, it gives the gated recurrent unit considerable performance in reducing the computational complexity of extremely long time series and improving the accuracy of prediction results. In addition, two open-source IEEE PHM Challenge and XJTU-SY bearing datasets sampled by the sensor are adopted to assess the prediction performance of the MHGRU with homology algorithm. Finally, some existing remaining useful life prediction approaches of rolling element bearings are used for comparison. The experimental results show that MHGRU is suitable for the remaining useful life prediction of rolling element bearings and is superior to other models.