Reducing false alarm rates in observer-based distributed fault detection schemes by analyzing moving averages

In this paper, we first analyze the possible limitations of a model-based fault detection method grounded on a partition-based distributed Luenberger observer. The corresponding fault detection test consists of comparing, for each time instant, the output prediction error with a suitable bound, computed analytically in a distributed and scalable way As a result, we highlight the presence of an often restrictive tradeoff between false-alarm and missed-detection rates. To overcome this significant drawback, we resort to a method based on the analysis of moving averages of residuals. Tests on an academic case study show the effectiveness of this approach.