关键词: 完工时间预测, 栈式自动编码器, 火箭总装, 精调网络, 预测精度

Abstract: To deal with the overdue problem caused by uncertainties and dynamic events during the rocket final assembly process,a novel approach based on Stacked Auto Encoders (SAE) was proposed.By substituting shallow Auto Encoder (AE) for traditional feature extraction process,the condensed non-linear feature of the cycle time's correlative factors was learned with unsupervised layer-wise pretraining.A fine-tuning network was constructed through stacking multiple shallow AE,and the cycle time prediction for rocket final assembly process was predicted accurately with supervised learning and parameter optimization process.Through the test data set generated from both simulation process and actual final assembly process in some spaceflight manufacture in Shanghai,the better generalization performance of proposed method was proved,and the prediction precision had improved compared with traditional prediction approaches.

Key words: cycle time prediction, stacked auto encoders, rocket final assembly, fine-tuning network, prediction precision