Knowledge integration strategies in defect diagnosis



Journal Title

Journal ISSN

Volume Title


Texas Tech University


Defect diagnosis—defined here as the process of evaluating and locating the true cause of a defect type—has been an island of automation and a time consuming and non-productive task. Needed are efficient and cost-effective methods which facilitate the task. The purpose of this research is to develop hybrid mathematical/simulation models and algorithms to diagnose defects with (I) multiple causes, (2) unknown cause probability, and (3) uncertain knowledge, with the objective of minimizing cost and number of trials.

The research problem is tackled in three phases. First, a diagnostic tree structure is proposed to (I) categorize diagnostic knowledge into sets of cause-effect relationships; and (2) simultaneously incorporate both testing costs and production loss. Then propositions for knowledge integration are developed to integrate initial and current knowledge, which correspond to the strength values for each edge within the diagnostic tree. Through the integration process, initial uncertain knowledge will be gradually pruned with newly arriving certain knowledge as the diagnosis task continues. Finally, primary elements of the conceptual decision process for troubleshooting defects are represented in a flow chart. Based on these ingredients, a linear multi-stage mathematic model is formulated, and a variety of knowledge integration strategies proposed.

Second, the problem of searching for the cause of a defect is formulated as a search problem where the estimated cause variable resembles a sensor function, and the true cause variable represents the target function. Therefore, the problem becomes a mapping of one function to the other. Several learning algorithms are created based on these developments. Then the algorithms are transformed into a probabilistic learning model where Monte-Carlo simulation is utilized to assess the performance of each algorithm. Primary propositions, lemmas and analytic properties are developed in this phase.

Third, a variety of experiments are used to investigate and compare the algorithms' (I) learning and fault-tolerant properties, (2) cost and trials performance, and (3) computational efficiency. Experimental results indicate that the proposed methods are superior to general techniques such as sequential and random searches in minimizing number of trials and costs. In addition, the proposed methods also contain learning and fault-tolerant properties.



Probability learning, Manufactures -- Defects