Bearing fault detection is very important to improve production efficiency and reduce accident rate in ball screw feed system. Deep learning-based analysis of bearing vibration data is a promising method for bearing fault detection. To make up for the shortcomings of traditional CNN and existing fault diagnosis algorithms, we construct a fault diagnostic method of ball screw bearings (2DCNN-LSTM-Attention) based on the integration of 2D Convolutional Neural Network (2DCNN), Long Short-Term Memory (LSTM) network and attention mechanism. Firstly, the 2DCNN is used to extract data features. Then, the time series features are processed by LSTM. Finally, the attention mechanism is used to amplify fault data features to improve diagnosis accuracy and efficiency. Combined with digital twin technology, a virtual-real mapping model is constructed to enhance the physical interpretability of feature expression. Comparative experiments of bearing fault diagnosis methods based on the Case Western Reserve University bearing dataset are carried out. The experimental and comparative results demonstrate that the model proposed in this study can effectively identify different fault types, and the average accuracy is 99.10%. Compared with the traditional 1DCNN, 2DCNN and 2DCNN-LSTM fault diagnosis methods, the proposed method improves both the diagnosis accuracy and efficiency.