Large and expensive mechanical equipment such as wind turbines generally has limited fault datasets from real-world operations for digital model development. This often leads to poor accuracy in implementing a model based on the life prediction. To address this data shortage issue in developing deep learning models, a remaining useful life prediction approach is proposed in this paper, which combines digital twin technology with transfer learning theory and the embedded convolutional long short-term memory (CLSTM) extended model. First, the main bearing of a direct-drive wind turbine is mapped to the digital world using digital twin technology, allowing the fault datasets of main bearings to be generated and thereby ensuring the model is trained sufficiently with a balanced dataset. The CLSTM network then performs convolutional operations on input-to-state and state-to-state transitions, thereby integrating the time dependence and time-frequency characteristics of the data. Meanwhile, transfer learning is used to transfer the trained model to the wind field for real-world fault diagnostics and the life prediction of the main bearings. Finally, the approach is applied to predict the life of the main bearings, and is also compared with other methods of similar type. The results verified that the proposed approach can effectively overcome the low data density of large equipment, greatly improving the accuracy of life prediction.