A framework for the self reconfiguration of automated visual inspection systems

Current automated visual inspection systems lack the flexibility demanded by the modern dynamic manufacturing environments in which the introduction and retirement of products is the norm. In these environments, it is difficult to reuse or reconfigure the inspection algorithms constructed using traditional approaches because of the considerable amount of time and effort required to adapt the inspection algorithms to perform the inspection of new products. This dissertation addresses this problem in two different ways: by proposing a structured framework for the design of efficient reconfigurable automated inspection systems and by developing methodologies for the solution of specific problems derived from this framework.; The first methodology aims to speed up the feature selection process during the algorithmic reconfiguration of the automated inspection systems. The methodology is based on the traditional stepwise variable selection procedure, but instead of using the conventional discriminant metrics such as Wilks-Lambda, it uses an estimation of the marginal classification error as the figure of merit to introduce new features into a quadratic classifier. This marginal error rate is estimated by using the densities of the conditional stochastic representations of the quadratic discriminant function. The application of the proposed methodology results in significant savings of computational time in the estimation of classification error over the traditional simulation and crossvalidation methods. Thus, the proposed methodology renders significant savings of time when reusing the preexisting inspection features to inspect the new products introduced into the assembly line.; The second methodology seeks to provide proactive design recommendations about the statistical characteristics of complementary features that minimize the total classification error when using with the preexisting features. The methodology is based on the conditional distributions of the quadratic classifier. The proposed methodology determines the values of the parameters of these distributions in the solution space determined by the canonical transformation of the original populations, and also provides a method to translate these values into the original populations' parameters. From the perspective of the development of auto reconfiguration systems, the proposed feature construction method has the potential of saving tremendous amount of time by directing the search for new features to particular domains.