Adaptive Methods for Fault Detection on Research Engine Test Beds

Abstract

In the case of test beds for research engines, fault detection methods that use models based on historical data face a particular challenge. Due to the experimental design of the test bed, offline training of statistical models with a data set containing all possible variations is simply not possible. The methods must adapt to the current data situation directly on-site. But this involves risks. First, computational time and memory requirements can become extremely large with high data volumes. Second, the data may be faulty and thus negatively affecting the models. To avoid both, a selection of data is made before it is used to build the fault-free reference model. For this purpose, a new statistic is presented as the combination of the Mahalanobis distance and the forecast residual. With it, it is possible to reduce the update frequency and to increase the rate of detected faulty points, since the models are no longer manipulated by faulty data points and thus the residuals provide a better structure for fault detection.