Extracting Information from Manufacturing Data using Data Mining Methods
Author: Giess, Matthew
Supervisor: Culley, Steve
Institution: University of Bath, Faculty of Engineering and Design
The manufacture and test of a product typically results in the generation of quantities of data describing both the characteristics of a product and its performance when tested. Such data contains information of potential use within design, and hence this research seeks to provide a means of both of extracting information from such data and making this information available for the design engineer. Various methods of data analysis exist and have been utilised for the analysis of manufacturing data. However, many such methods rely upon generating data specifically for analysis, which has cost implications, or simply quantify a previously defined relationship. Others, such as Taguchi’s Robust Design, require the ability to control the values of variables, limiting its application when the variance of characteristics within manufacturing tolerance is under investigation. The field of Data Mining (DM) allows for data to be analysed and modelled without the need to specify expected relationships, and without the need to control the value of variables. In a number of DM implementations within engineering it was noted that little attention was given to the methods by which data were generated, and also how useful information might be extracted from the resultant DM models and subsequently validated. This thesis seeks to address both of these areas.
Information Extraction and Validation
An analytical model was used to generated data mimicking that typically collated during a manufacturing process, and DM modelling methods were used to analyse these data. Various methods of information extraction were used, of which some were novel, and the extracted information was compared against industrially-validated information obtained from the analytical model itself. Encouraging results were obtained.
Data Generation within Manufacturing
Two industrial case studies were used to assist in understanding the nature of manufacturing data. Methods of retrospectively removing erroneous data were tested. Numerous data generation issues were noted, and these examples were used to create a three-tier hierarchy which guides consideration of data generation practices.