Deflection yoke (DY) is one of the core components of the color display tube (CDT) in a computer monitor that determines the image quality. Once a DY anomaly is found, the remedy process consists of two steps: identifying predefined symptoms from an anomalous display pattern and adjusting manufacturing process parameters. This study focuses on eliciting expert diagnostic strategies for the identification of DY symptoms by applying systematic and quantitative data processing. The methods of stepwise regression, inductive learning, and expert interviewing are integrated into the knowledge acquisition process. The regression model is used to select the attributes of display patterns that are of the most stochastic significance for symptom identification. Inductive learning with the selected attributes is then performed to generate decision trees from the accumulated training data. The trees are transformed to rule sets which can be verified by their prediction of DY symptoms from evaluation data. The incorrectly predicted cases thus obtained are examined in the interviews with the experts and refined. The knowledge base is thus rectified and augmented to undergo the same process again until the performance criteria are satisfied. In this case, only after two revisions of the knowledge base, the system performed as well as the process engineers in the evaluation with raw field data. The knowledge eliciting and refining process proved to be effective and repeatable to achieve continuous refinement of the diagnostic performance and better outgoing quality of the product.