Information-Theoretic Syndrome and Root-Cause Evaluation

Abstract

Reasoning-based methods are promising because a detailed system model is not needed to construct the diagnosis system O’Farrill et al. (Optimized reasoning-based diagnosis for non-random, board-level, production defects, (2005), [1]), Fenton et al. (IEEE Trans Syst Man Cybern Part C: Appl Rev 31:269–281, [2]). The diagnosis engine is incrementally built based on the database of successfully repaired faulty boards. Machine learning techniques enable reasoning-based diagnosis, leveraging ease of implementation, high diagnosis accuracy, and continuous learning. Repairs of faulty components are suggested through a ranked list of suspect components, e.g., based on artificial neural networks (ANNs) and support-vector machines (SVMs) [3]. However, a machine learning-based diagnosis system requires a sizeable database for training. Ambiguous root-cause identification may result if the diagnosis system lacks a sufficient database with adequate information for mapping errors to root causes. There is a need for a rigorous framework that can evaluate an existing diagnosis system using quantifiable metrics, and provide means for enhancing the accuracy of diagnosis. Ambiguity in root-cause identification confounds diagnosis, hence such ambiguities need to be identified and used as a guideline for test development. The accuracy of diagnosis and time needed for diagnosis also depend on the quality of syndromes (erroneous observations). Redundant or irrelevant syndromes not only lead to long diagnosis time, but also increase diagnosis complexity.