Acquiring diagnostic knowledge from documents to predict issues in aircraft assembly
MetadataShow full item record
Expert knowledge is important in a product's lifecycle, especially during the manufacturing part of the lifecycle. Most of such knowledge is obtained through experience and its reuse can help prevent potential issues in subsequent product development. Extracting the knowledge acquired during one development cycle for reuse in subsequent development closes the knowledge loop within a product's lifecycle. This thesis is aimed at acquiring expert knowledge from the manufacturing and assembly phases in aerospace manufacturing for eventual use in the planning and design stages. Given the difficulties in knowledge acquisition from experts, this thesis focuses on acquiring expert knowledge from text documents. The proposed method consists of three parts - segregation of relevant text, acquisition of issues, causes and parameters, and realising context of knowledge. These parts become the research questions to be addressed in this thesis. The segregation of relevant text involves identifying coherent segments and then classifying the relevant segments. The method proposed for segregation is based on discourse representation that treats documents as a discourse, and attempts to measure topic changes by looking at the relatedness between the discourse entities. A measure for calculating similarity between sentences is used for identifying segments. Implementation and validation of the method with human subjects is described. The acquisition of diagnostic knowledge is performed by first identifying the parts of text containing issues. Functional modelling and sentiment analysis are considered, and the latter is chosen. A strategy for finding the causes of issues based on text patterns and sentiment is proposed. Once the issues and causes are known, the relations and parameters that form the cause are dissected to represent the knowledge as rules in a knowledge base. A hybrid method using both text patterns and dependency parse of the text representing cause is proposed. The knowledge acquisition pipeline from the issues and causes to the rules is implemented. The acquisition of causes and effects is also cross-validated with human subjects. Once the knowledge is acquired, methods for capturing the context (in which it was acquired or it needs to be applied) are proposed. Context containers that make use of proximity of words to ve factors defined to influence assemblability are used to capture the context. It is expected that assembly situations can be described using these, and enable to match the knowledge to a similar situation when reusing the knowledge. At the application stage, an assembly situation model is proposed that combines product and process information to model the application situation. The application of knowledge and implementation of these models is part of a legacy-knowledge based smart manufacturing system under development. This thesis concludes with a discussion of how the objectives were met, how the current methods could be improved and the directions in which the methods could be extended.