Purpose <p>The life prediction of rolling bearings is an important part of the field of predictive maintenance, and the prediction of its stability and reliability is crucial. This process requires data-driven models that not only exhibit a high degree of predictive accuracy but also have broad applicability. However, when dealing with big data, we often encounter problems of difficulty in data extraction and insufficient data volumes, which further leads to the inadequacy of a single deep learning model in learning small sample data and the loss of feature extraction.</p> Methods <p>Research shows that Capsule Networks can use capsule and dynamic routing mechanisms to extract deep structural features in data, while BiGRU performs well in time series prediction and can capture both historical and future feature information. In view of this, this study proposes a new bearing remaining useful life (RUL) prediction framework (CN-BiGRU) combined Capsule Networks and multi-layer BiGRU. This framework alternately stacks Capsule Networks and multi-layer BiGRU, using the original vibration signal as the data input source for feature extraction and RUL prediction.</p> Results <p>The work of this research has the following salient features: (1) The specially designed effective combination network of BiGRU layer and capsule network enhances the model's learning ability for time series by connecting the output and input residuals at each step. (2) Multi-level residual connections are used at each stage of the model to promote gradient flow, thus improving the generalization ability and stability of the model. This study uses two public experimental data (IEEE PHM 2012 and XJTU-SY) to experimentally verify the model.</p> Conclusion <p>Experimental results show that CN-BiGRU also shows good prediction performance on small sample data, which is better than the existing mainstream prediction methods and provides a new solution for rolling bearing RUL prediction.</p>