Decision Fusion Methods With Applications to Structural Health Monitoring

Although low-level data and information fusion process improves the accuracy and robustness of decision-making about events, there is a clear justification and need for decision fusion at the highest processing level if integrated health monitoring systems are of interest. Several decisions at low-level of abstraction may be produced by different decision-makers (classifiers), however, fusion at the decision’s level is required at the final stage of the process in order to provide accurate assessment about the monitored system as a whole. An example of such integrated systems with complex decision-making scenarios is the structural health monitoring (SHM) of aircrafts. Several decisions can be produced by different diagnostics algorithms monitoring different components of the aircraft structure. To better understand and assess events happening in the aircraft, decisions made by these diagnostic algorithms must be fused to provide an accurate picture about the true health of the aircraft. Ideally, such high-level decision fusion reduces the level of uncertainty in the decisions made by the monitoring system leading to more trusted decisions with high-level of confidence. Profound understanding of the characteristics of the decision fusion methodologies and algorithms reported in the literature is a crucial step for successful implementation and practical realization of the above mentioned process of decision fusions. In this paper we will present two topics of importance to the decision-making and decision fusion algorithms; their characteristics and relative performance. First, we will provide brief reviews of the major decision fusion methods reported in the literature and group them into classes. Then, we will provide brief reviews of its theoretical basis and conducted analysis of its performances. Second, candidate algorithms belonging to different fusion methodologies will be selected and its performance based on synthetic decisions produced with various interaction complexities will be tested. We made sure that the used synthetic decisions are unbiased and provide a level playing field for fair performance comparison of the selected decision fusion algorithms. Finally, the performance of the most promising candidates of the fusion algorithms will be verified on information extracted from experimental data collected from structural health monitoring experiments. A comprehensive analysis and testing results will be reported in the paper.