Anomaly Detection in Large-Scale Wind Tunnel Tests Using Gaussian Processes

摘要

We propose a method to monitor and analyze correlations between sensors in multiple wind tunnel measurement systems using a novel distance metric within a Gaussian process framework. This method allows us to predict an individual sensor's output by considering the joint output of multiple sensor systems measuring different physical quantities over the course of several experimental trials. We use this method to detect and potentially correct aberrant sensor readings. We illustrate the method using data from five distinct sensor systems, collected during a three week experimental campaign in the Virginia Tech Stability Wind Tunnel.