Computing similarity of text-based assembly processes for knowledge retrieval and reuse

• Four sentence similarity methods are used for analysis of assembly work instructions.
• A survey was conducted to compare computational sentence similarity to human interpretation of assembly work instruction similarity.
• Sensitivity of all four methods to use of synonyms was assessed.

The objective of this research is to use text similarity algorithms to enable retrieval and knowledge sharing of text-based assembly process plans. In distributed design-manufacturing enterprises, there exists a need to: (1) establish and use “best practice” assembly process descriptions to ensure process design consistency across manufacturing locations and (2) leverage and adapt existing process design knowledge across product lines. In this research, the need for better communication is addressed by design knowledge reuse. Specifically, previously-authored assembly processes are retrieved from a centralized repository with a text-based similarity and retrieval algorithm. The similarity of forty-five text-based assembly work instruction pairs (obtained from ten work instruction sets) is computed using four text mining algorithms: (1) Word Overlap, (2) Jaccard Score, (3) TF–IDF and (4) Latent Semantic Analysis. The similarity scores are used to compare and retrieve similar work instructions from a repository of existing assembly processes descriptions. A survey is conducted to develop a baseline quantification of assembly work instruction similarity. The scores from each text comparison method are compared to the scores from the survey. A statistical hypothesis test shows the Jaccard method mimics human interpretation of assembly work instruction similarity better than the three other text comparison methods. However, the Jaccard method is insensitive to synonymy and polysemy of words. The Latent Semantic Analysis method is relatively insensitive to synonymy and polysemy of words; and was found to have a difference of 0.1 with respect to survey data. This indicates that Latent Semantic Analysis can be used to retrieve assembly work instructions authored in free text. By doing so, engineers will be presented with similar variants of assembly work instructions have been authored. This will allow engineers to compare and assess the efficiency of their assembly process and gain insight into how other facilities are performing similar assembly operations.

Average correlation to human interpretation of assembly work instruction similarity. The correlation between manufacturing human experts and different text mining algorithms are evaluated to determine how computational algorithms can be used to efficiently mine manufacturing big data to enable reuse.Figure optionsDownload as PowerPoint slide