<p>Gyro drift tested in the assembling process of dynamic gyros is the key indicator affecting their guidance accuracy. In traditional drift debugging, the selection of drift-adjusting screw still relies on the trial and error of screw mass, resulting in low debugging accuracy and high repetition rate. To tackle this problem, we propose a VGAE-SGAE model to predict the drift-adjusting screw mass in high accuracy. In this paper, graph neural network is adopted for interactive feature extraction of gyro assembly parameters. A VGAE model is designed for assembly graph data augmentation under the small-sample condition. Generated and original assembly data is utilized for the training of an SGAE model to predict the drift-adjusting screw mass. The experimental results show that the proposed method can achieve drift-adjusting screw mass prediction and outperforms all other existing methods,lowering the debugging repetition rate and improving gyro assembly efficiency.</p>