Abstract: Aiming at the poor applicability of existing process generation methods to different professions,a process generation method based on deep learning encoder-decoder structure was proposed.The characteristics of process attributes were extracted from the process outline file through the encoder,and the text representation vector of process attributes was formed.The decoder generated the process gradually according to the representation vector.Based on the data of two professional process outline files of sheet metal parts manufacturing and assembly,24 kinds of encoder-decoder structures were compared and studied,the highest accuracy rates on the two data sets were 0.8287 and 0.6973 respectively,which meant that 82.87% and 69.73% of the generated process were directly accepted without subsequent modification.On the one hand,it showed that the proposed method could effectively learn the relationship between process attributes and process methods from data to generate the process.On the other hand,the same encoder-decoder structure was used in both specialties,indicating the applicability of the proposed method for different specialties with migration capability.

Key words: process generation, encoder-decoder, deep learning, manufacturing and assembly

摘要： 针对现有工艺过程生成方法对于不同专业适用性较差的问题,提出一种基于深度学习编码器—解码器结构的工艺过程生成方法。该方法利用工艺大纲文件数据,通过编码器提取大纲文件中工艺属性的特征,形成工艺属性文本表征向量,解码器根据表征向量逐步生成工艺过程。在钣金零件制造与装配两个专业工艺大纲文件数据上,比较研究了24种编码器—解码器结构,最高准确率分别达到0.8287和0.6973,即生成的工艺过程有82.87%和69.73%可直接接受,不需要后续修改。这一方面表明所提出方法能够有效地从数据中学习工艺属性与工艺方法间的关系,从而生成工艺过程;另一方面,在两个专业采用相同编码器—解码器结构,表明所提方法对于不同专业的适用性,具有迁移能力。

关键词: 工艺过程生成, 编码器—解码器, 深度学习, 制造与装配