Using Bayes belief networks in industrial FMEA modeling and analysis

This paper presents the use of Bayes probabilistic networks as a new methodology for encoding design failure modes and effects analysis (BN-FMEA) models of mechatronic systems. The method employs established Bayesian belief network theory to construct probabilistic directed acyclic graph (DAG) models which represent causal and statistical dependencies between system-internal and -external (customer and world) state and event variables of the physical system. A new class of severity variables is also defined. Root probabilities and conditional probability and severity utility tables are generated and attached to the graph structure for use in inferencing and design trade-off evaluation. BN-FMEA provides a language for design teams to articulate-with greater precision and consistency and less ambiguity-physical system failure cause-effect relationships, and the uncertainty about their impact on customers and the world. Demonstration software developed at Stanford illustrates how BN-FMEA can be applied to FMEA modeling of an inkjet printer. The software supports knowledge acquisition of BN-FMEA models, and generates from the belief net model criticality matrices and Pareto charts conformant with established FMEA standards such as SAE 1998. The approach supports traditional design FMEA objectives, identification of system failure modes, and provides improved knowledge representation and inferencing power. Limitations of the BN-FMEA methodology are also discussed. Finally, BN-FMEB is presented as a basis for improved integration of design and diagnostic modeling of mechatronic systems