|
Abstract:
|
Wind tunnels are intricate networks of sensors used to learn how different test apparatuses, such as airfoils, will react in environmentally controlled spaces. Thus, large-scale multi-type sensor networks require continual monitoring to ensure the quality of the test apparatus and to reduce the amount of corrupted data, which leads to increase experimental costs, wasted time, and spurious engineering conclusions. Previous monitoring methodologies for detecting anomalies in wind tunnels included Gaussian Processes, Principal Component Analysis, and robust Bayesian regressions. However, our experimental setup involves continual instrumental movement, wind tunnel re-configurations, or low-levels of adverse observations in training datasets that limit these methodologies. In this paper, we propose a mixture model of signals to account for cases where sensors are intermittently collecting anomalous signals. To demonstrate the utility of separating signal components, we compare two Bayesian monitoring methodologies: (1) a Bayesian, heavy-tailed regression, and (2) our proposed Normal-Cauchy mixture model with Principal Component Analysis through simulation and real-world case studies.
|
