Knowledge Modeling Method for Assembly Process Planning based on Process History

Assembly processes account for over 50 % of production time and are still largely manual in highly variable industries. Current planning is based on expert knowledge and previous planning results, leading to inefficiencies with new workers. This study presents a deep learning model that uses historical process data to support similarity- and variant-based planning. The model predicts subsequent process steps and components based on a given process sequence by training a Bidirectional encoder representations from transformers (BERT)-based architecture on 280 different work instructions, achieving a combined accuracy of 94 % and an F1-score of 82%. The results show the potential of assisted planning in creating work instructions. However, it is impossible to generate work instructions for the planning of completely new products..