Failure modes and effects analysis with Bayesian belief networks: Bridging the design-diagnosis modeling gap

Product design and diagnosis are today worlds apart. Despite strong areas of overlap at the ontological level, traditional design process theory and practice do not recognize diagnosis as a part of the design modeling or product engineering process chain; neither do diagnosis knowledge engineering processes employ design phase models as explicit sources of knowledge acquisition for diagnostic models. Most design phase models, such as conventional failure modes and effects analysis (FMEA) models, lack sufficient knowledge and structure so as to be useful for direct conversion to diagnostic models; such models typically lack critical knowledge of functions and causal relations needed for diagnostic inferencing.; This dissertation presents a new methodology for encoding design FMEA models in a manner intended to bridge the design-diagnosis modeling gap. The approach—Bayesian network FMEA (BN-FMEA)—employs Bayesian belief networks to represent product failure causal models, and to provide such models with adequate structure so as to support the use of formal reasoning methods. We define and introduce a new class of conditional severity variable; severity nodes are attached to failure scenario chains in the graph in order to represent the severity distribution of failure end-events. The methodology employs standard Bayesian inference algorithms to obtain failure occurrence probabilities and severity distributions from the belief network. These are then used to build the standard FMEA Criticality Matrix. The approach is demonstrated with the aid of practical examples taken from the FMEA analysis of a commercial inkjet printer. Such Bayesian network-based design FMEA models can be edited to create diagnostic models, thereby enabling reuse of design phase knowledge.; Secondly, the dissertation adopts an explicit enterprise-level knowledge engineering perspective to the formalized integration of design and diagnosis tasks and models. We propose and implement parts of the DAEDALUS knowledge engineering framework, in which product engineering tasks and models are integrated around a common domain ontology and product models library. The approach uses the ontology to populate BN-FMEA and diagnosis models with relevant concepts (variables) and concept relations, thus serving as a shared concept dictionary-style mechanism for knowledge sharing and reuse across different design and diagnosis modeling environments.