Functional Anomaly Detection: a Benchmark Study

  • 2022-01-13 18:20:32
  • Guillaume Staerman, Eric Adjakossa, Pavlo Mozharovskyi, Vera Hofer, Jayant Sen Gupta, Stephan Clémençon
  • 2


The increasing automation in many areas of the Industry expressly demands todesign efficient machine-learning solutions for the detection of abnormalevents. With the ubiquitous deployment of sensors monitoring nearlycontinuously the health of complex infrastructures, anomaly detection can nowrely on measurements sampled at a very high frequency, providing a very richrepresentation of the phenomenon under surveillance. In order to exploit fullythe information thus collected, the observations cannot be treated asmultivariate data anymore and a functional analysis approach is required. It isthe purpose of this paper to investigate the performance of recent techniquesfor anomaly detection in the functional setup on real datasets. After anoverview of the state-of-the-art and a visual-descriptive study, a variety ofanomaly detection methods are compared. While taxonomies of abnormalities (e.g.shape, location) in the functional setup are documented in the literature,assigning a specific type to the identified anomalies appears to be achallenging task. Thus, strengths and weaknesses of the existing approaches arebenchmarked in view of these highlighted types in a simulation study. Anomalydetection methods are next evaluated on two datasets, related to the monitoringof helicopters in flight and to the spectrometry of construction materialsnamely. The benchmark analysis is concluded by recommendation guidance forpractitioners.


Quick Read (beta)

loading the full paper ...