Cosmetic quality of surfaces: a computational approach

Tools

Balendran, V., 1993. Cosmetic quality of surfaces: a computational approach. PhD, Nottingham Trent University.

Preview

Text
10290074.pdf - Published version
Download (21MB) | Preview

Official URL: http://gateway.proquest.com/openurl?url_ver=Z39.88...

Abstract

The use of qualitative, subjective, personal experiences for decision making purposes in manufacturing processes is examined. The dissertation offers directions for mechanising such experiences and decision making processes by taking an example from the automotive industry - the assessment of cosmetic quality of vehicle body panels - which has defied automation and still remains as an enclave of human activity within the industry. Cosmetic quality is essentially the assessment of the degree of eye appeal of body panels, taking into account the various defects on its surface, which are visible to the trained eye of inspectors under special lighting conditions. The difficulty in dealing with the problem arises due to the non-verbalizability of both the sensory content of the inspection experience, and the decision making process that follows. This is identified as due to limitations of language as well as the impossibility of gaining awareness of some of the thought processes that are involved.

The general solution offered requires the mapping of the human experience - both the sensory and the decision making - to mathematical domain with an appropriate basis. Specifically it requires; 1. a new tool and/or a methodology for mapping of the sensory content of the experience, 2. a new mathematical basis for evaluation of the experience that enables hitherto unknown forms of knowledge to be synthesised and acquired, and 3. the employment of conventional techniques in Artificial Intelligence (especially Expert systems. Machine Learning and Neural Networks), with the new knowledge as data, for the mapping of the decision making process.

With regard to cosmetic quality a novel computer vision based mechanical tool has been designed and implemented, which maps the visual experience of inspection as a mathematical "cosmetic map". A new mathematical basis is proposed for the evaluation of the cosmetic map, which results in the assignment of a descriptive classification code for any panel inspected by the tool. When this machine classification and actual human classification are taken as a pair, it embodies new inspection knowledge, and is acquired on-line in a database. While the potential of this evolving database to function as an "expert assistant" is exploited, its inability to reach generalisations is also indicated as a disadvantage. After sufficient inspection data has been acquired in such a database (over a period of time), employing machine learning techniques for unravelling the rules involved with decision making is highlighted as a way of overcoming this disadvantage. However, in the absence of such representative data, as is the case here, how synthetic data can be composed and used to determine experimentally a suitable neural network for generalising its learning and thereby functioning as an expert assistant is also indicated. The internal representation learnt by the selected network seemed reasonable, if not human-like, and strengthens the view that networks have the power to capture salient features or concepts, not explicitly stated in the data.

Item Type:

Thesis

Creators:

Balendran, V.

Date:

1993

ISBN:

9781369323238

Identifiers: